<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
  <channel>
    <title>BroadRouter blog</title>
    <link>https://broadrouter.com/blog/</link>
    <atom:link href="https://broadrouter.com/blog/rss.xml" rel="self" type="application/rss+xml" />
    <description>LLM API pricing, agents, Arabic AI and data retention, from the BroadNet Inference team.</description>
    <language>en</language>
    <item>
      <title>LLM API price changes in October 2026: 3 rises, no cuts</title>
      <link>https://broadrouter.com/blog/llm-api-price-changes-october-2026/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/llm-api-price-changes-october-2026/</guid>
      <pubDate>Mon, 16 Nov 2026 08:00:00 GMT</pubDate>
      <category>Pricing</category>
      <description>Few, and no cuts. Comparing archived pricing pages of 7 LLM API providers from early October to early November 2026, we found 3 list-price rises: two at DeepInfra (Mistral Nemo, Gemma 4 31B) and one at Together AI, where Qwen3.8 Flash went back to its September price. 9 models were newly priced and 11 left the pages.</description>
    </item>
    <item>
      <title>A private endpoint starts with a clear workload</title>
      <link>https://broadrouter.com/blog/private-endpoint-planning/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/private-endpoint-planning/</guid>
      <pubDate>Sun, 15 Nov 2026 08:00:00 GMT</pubDate>
      <category>Engineering</category>
      <description>Before you ask for a private endpoint, write down the workload: requests per day and at peak, tokens in and out, the concurrency that implies, the context you need, the processing region, the model and its licence, and a budget. Agree region and limits in writing, then smoke-test the issued endpoint before real data.</description>
    </item>
    <item>
      <title>Add diacritics to Arabic text with an LLM: prompt, checks, DER</title>
      <link>https://broadrouter.com/blog/arabic-diacritization-llm/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/arabic-diacritization-llm/</guid>
      <pubDate>Sun, 15 Nov 2026 08:00:00 GMT</pubDate>
      <category>Arabic AI</category>
      <description>Ask the model for full tashkeel and nothing else, then strip the diacritics from its reply and compare with your input. Any changed letter means reject and retry. Score a hand-checked sample with diacritic error rate (DER) and word error rate (WER), with and without case endings. Published results put GPT-4 close to, but behind, the best dedicated diacritizers.</description>
    </item>
    <item>
      <title>Arabic grammar and spelling correction with an LLM: a how-to</title>
      <link>https://broadrouter.com/blog/arabic-grammar-spelling-correction-llm/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/arabic-grammar-spelling-correction-llm/</guid>
      <pubDate>Sun, 15 Nov 2026 08:00:00 GMT</pubDate>
      <category>Arabic AI</category>
      <description>Ask the model for the corrected text only, with the fewest possible edits and no rephrasing. Then diff the reply against the input in code: reject replies that changed too much, and label each word-level edit as hamza, taa marbuta, alef maqsura, merge, split or other. Score a hand-corrected sample on edits, with precision, recall and F0.5.</description>
    </item>
    <item>
      <title>Arabic meeting minutes from recordings: an API pipeline</title>
      <link>https://broadrouter.com/blog/arabic-meeting-minutes-from-recordings/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/arabic-meeting-minutes-from-recordings/</guid>
      <pubDate>Sun, 15 Nov 2026 08:00:00 GMT</pubDate>
      <category>Arabic AI</category>
      <description>To make Arabic meeting minutes with an API, transcribe the recording with an Arabic speech-to-text model, split long transcripts into sections, ask an LLM for decisions, action items and open questions as JSON, then merge the sections. Have the model quote the transcript for every action item, and write the minutes in Modern Standard Arabic or English.</description>
    </item>
    <item>
      <title>Arabic named entity extraction with an LLM: JSON, dates and checks</title>
      <link>https://broadrouter.com/blog/arabic-named-entity-extraction-llm/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/arabic-named-entity-extraction-llm/</guid>
      <pubDate>Sun, 15 Nov 2026 08:00:00 GMT</pubDate>
      <category>Arabic AI</category>
      <description>Ask the model to copy each entity exactly as written, with a type from a fixed list, as JSON. Check every entity in code: it must appear in the source once diacritics are stripped and alef forms unified. Convert Hijri dates with a library such as hijridate, not the model. Score a hand-labelled sample with entity-level precision, recall and F1.</description>
    </item>
    <item>
      <title>Arabic OCR API comparison: Google, Azure, Textract, Mistral, LLMs</title>
      <link>https://broadrouter.com/blog/arabic-ocr-api-comparison/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/arabic-ocr-api-comparison/</guid>
      <pubDate>Sun, 15 Nov 2026 08:00:00 GMT</pubDate>
      <category>Arabic AI</category>
      <description>Four OCR APIs list Arabic printed text in their own docs: Google Cloud Vision, Azure AI Document Intelligence, Amazon Textract and Mistral OCR. Google, Azure Read and Textract each charge USD 1.50 per 1,000 pages at entry volume. Only Azure lists Arabic handwriting. Vision LLMs, such as Qwen3.8 27B on BroadNet, bill a page as tokens and return the fields you ask for.</description>
    </item>
    <item>
      <title>Arabic subtitles (SRT) from a speech-to-text API: what you need</title>
      <link>https://broadrouter.com/blog/arabic-subtitles-srt-api/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/arabic-subtitles-srt-api/</guid>
      <pubDate>Sun, 15 Nov 2026 08:00:00 GMT</pubDate>
      <category>Arabic AI</category>
      <description>An SRT file needs text plus a start and end time for every caption. APIs with word or segment timestamps, or direct SRT export, give the best timing. BroadNet Transcribe Arabic returns text only, so split audio at pauses, transcribe each chunk and time captions from the chunk offsets. Fine for talks and podcasts, not for frame-accurate film subtitles.</description>
    </item>
    <item>
      <title>Arabic text classification with an LLM: sentiment, intent, triage</title>
      <link>https://broadrouter.com/blog/arabic-text-classification-llm/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/arabic-text-classification-llm/</guid>
      <pubDate>Sun, 15 Nov 2026 08:00:00 GMT</pubDate>
      <category>Arabic AI</category>
      <description>To classify Arabic text with an LLM, define a short label set with one-line definitions, give a few labelled examples in each dialect you see, ask for JSON with one label and a short quote as evidence, and set reasoning low. Then score the model on 200 hand-labelled messages with macro F1 before trusting it. Fine-tuned Arabic models can still beat a general LLM.</description>
    </item>
    <item>
      <title>Arabic text normalization for search and RAG: what to change, what not</title>
      <link>https://broadrouter.com/blog/arabic-text-normalization-search/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/arabic-text-normalization-search/</guid>
      <pubDate>Sun, 15 Nov 2026 08:00:00 GMT</pubDate>
      <category>Arabic AI</category>
      <description>For search and RAG indexes, normalize both the stored copy and every query the same way: Unicode NFKC, strip tashkeel and tatweel, unify alef forms, map ى to ي and ة to ه, and convert digits. Keep the original text for display and for the LLM. Treat hamza seats and ta marbuta as tunable: they raise recall but merge some distinct words.</description>
    </item>
    <item>
      <title>Arabic text summarization with an LLM: prompts, length and scoring</title>
      <link>https://broadrouter.com/blog/arabic-text-summarization-llm/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/arabic-text-summarization-llm/</guid>
      <pubDate>Sun, 15 Nov 2026 08:00:00 GMT</pubDate>
      <category>Arabic AI</category>
      <description>Name the register (usually Modern Standard Arabic, even for dialect input) and a word range in the prompt. Then check in code: count the words, confirm every number in the summary appears in the source, and retry once with the problem stated. Score with an Arabic-aware ROUGE tokenizer plus human review; default rouge-score gives Arabic 0.</description>
    </item>
    <item>
      <title>Artificial Analysis alternatives: where to compare LLMs in 2026</title>
      <link>https://broadrouter.com/blog/artificial-analysis-alternatives/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/artificial-analysis-alternatives/</guid>
      <pubDate>Sun, 15 Nov 2026 08:00:00 GMT</pubDate>
      <category>Comparisons</category>
      <description>Artificial Analysis is the main independent source for LLM speed, latency and intelligence scores per API provider. Each alternative measures something else: Arena ranks human preference, Epoch AI and Vals AI publish benchmark results, OpenRouter ranks usage, and price databases such as BroadRouter's (BroadNet's site) list prices only. Use two side by side, then test on your own data.</description>
    </item>
    <item>
      <title>Before fine-tuning: define the problem to solve</title>
      <link>https://broadrouter.com/blog/fine-tuning-readiness/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/fine-tuning-readiness/</guid>
      <pubDate>Sun, 15 Nov 2026 08:00:00 GMT</pubDate>
      <category>Engineering</category>
      <description>Fine-tune only after an evaluation shows a recurring failure that better prompts, examples, retrieval and validation did not fix. Missing knowledge usually needs retrieval, and format errors need structured output and checks. Before asking for a custom model, collect labelled examples you may legally use, hold out a test set, and record the score to beat.</description>
    </item>
    <item>
      <title>BroadNet's GPU hardware and serving approach, as BroadNet describes it</title>
      <link>https://broadrouter.com/blog/broadnet-gpu-fleet-and-serving/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/broadnet-gpu-fleet-and-serving/</guid>
      <pubDate>Sun, 15 Nov 2026 08:00:00 GMT</pubDate>
      <category>Engineering</category>
      <description>BroadNet says it runs one serving stack on three hardware tiers: data-center GPUs from A100 to B300, a workstation tier of RTX and DGX Spark cards that carries the always-on public API, and appliances on client sites. It reports about 270 models run across 13 GPU families, NVIDIA and AMD. Some API models are served by US inference partners instead.</description>
    </item>
    <item>
      <title>Cache LLM API responses to cut cost: an exact-match cache in Python</title>
      <link>https://broadrouter.com/blog/cache-llm-responses-cut-cost/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/cache-llm-responses-cut-cost/</guid>
      <pubDate>Sun, 15 Nov 2026 08:00:00 GMT</pubDate>
      <category>Engineering</category>
      <description>Hash the full request (model, messages, temperature, max_tokens, response_format, tools, reasoning settings and your prompt version) into a key, and store the answer under it with an expiry. Serve a stored answer only for deterministic, side-effect-free requests. Measure how often identical requests repeat in your logs first: that share is the most an exact cache can save.</description>
    </item>
    <item>
      <title>Claude Code with open models: proxies, native endpoints and limits</title>
      <link>https://broadrouter.com/blog/claude-code-with-openai-compatible-models/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/claude-code-with-openai-compatible-models/</guid>
      <pubDate>Sun, 15 Nov 2026 08:00:00 GMT</pubDate>
      <category>Agents</category>
      <description>Claude Code sends Anthropic Messages API requests to ANTHROPIC_BASE_URL. To use an OpenAI-compatible provider, run a translating proxy such as LiteLLM and point Claude Code at it, or pick a provider with its own Anthropic-format endpoint, such as DeepSeek or Z.ai. Anthropic does not support non-Claude models in Claude Code, so test tools, caching and context limits yourself.</description>
    </item>
    <item>
      <title>Design an AI experience for multiple languages</title>
      <link>https://broadrouter.com/blog/multilingual-product-design/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/multilingual-product-design/</guid>
      <pubDate>Sun, 15 Nov 2026 08:00:00 GMT</pubDate>
      <category>Arabic AI</category>
      <description>Store two settings: the interface locale and the content language, and never infer one from the other. Set direction from the locale, isolate user and model text with dir="auto" or bdi, format numbers and dates with the locale, normalise digits on input, treat language detection as a hint, and evaluate quality and cost per language.</description>
    </item>
    <item>
      <title>Do LLM API prices only go down? 52 rises and how to budget for them</title>
      <link>https://broadrouter.com/blog/llm-api-price-increases/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/llm-api-price-increases/</guid>
      <pubDate>Sun, 15 Nov 2026 08:00:00 GMT</pubDate>
      <category>Pricing</category>
      <description>No. In archived pricing pages from 11 providers, we found 52 list-price rises on the same model between 2024 and November 2026, against 100 cuts. The median rise was 40% on a blended price, and 16 rises doubled it or more. Most were at DeepInfra (40) and Together AI (7), but Anthropic, Mistral and DeepSeek raised prices too.</description>
    </item>
    <item>
      <title>Egypt's data protection law (Law 151 of 2020) and LLM APIs</title>
      <link>https://broadrouter.com/blog/egypt-data-protection-law-llm-api/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/egypt-data-protection-law-llm-api/</guid>
      <pubDate>Sun, 15 Nov 2026 08:00:00 GMT</pubDate>
      <category>Privacy</category>
      <description>Under Egypt's Personal Data Protection Law (Law 151 of 2020), a prompt with personal data sent to an LLM API abroad is a cross-border transfer. Article 14 allows it only if the destination protects data at least as well as Egyptian law and the Personal Data Protection Center licenses it. Decision 816 of 2025 adds consent and filing rules. Not legal advice.</description>
    </item>
    <item>
      <title>How often do LLM API prices change? 3 years of archived pricing pages</title>
      <link>https://broadrouter.com/blog/how-often-llm-api-prices-change/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/how-often-llm-api-prices-change/</guid>
      <pubDate>Sun, 15 Nov 2026 08:00:00 GMT</pubDate>
      <category>Pricing</category>
      <description>Rarely, for most models. In 308 archived pricing pages from 11 providers (January 2023 to November 2026), 398 of 498 models seen in more than one month (80%) never changed price. We found 161 list-price changes: 100 cuts, 52 rises and 9 mixed. DeepInfra made 121 of them. OpenAI, Anthropic and Google repriced 8 of 126 models between them.</description>
    </item>
    <item>
      <title>How to fix invalid JSON from an LLM: causes, repair, retry</title>
      <link>https://broadrouter.com/blog/fix-invalid-json-from-llm/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/fix-invalid-json-from-llm/</guid>
      <pubDate>Sun, 15 Nov 2026 08:00:00 GMT</pubDate>
      <category>Engineering</category>
      <description>Check finish_reason first: a reply cut off at max_tokens cannot be repaired safely, so raise the limit and resend. Otherwise strip code fences and surrounding prose, decode the first JSON value, then fix only unambiguous mistakes such as trailing commas. Validate against a schema, and if that fails, send the exact error back once or twice. Structured output prevents most of this where supported.</description>
    </item>
    <item>
      <title>How to reduce LLM API costs: a 10-step checklist, in order</title>
      <link>https://broadrouter.com/blog/how-to-reduce-llm-api-costs/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/how-to-reduce-llm-api-costs/</guid>
      <pubDate>Sun, 15 Nov 2026 08:00:00 GMT</pubDate>
      <category>Pricing</category>
      <description>To reduce LLM API costs, first log tokens per request, then work in this order: keep the prompt prefix stable so it caches, trim prompts, cap output and reasoning, use the smallest model that passes your tests, batch or shift work that can wait, cache repeated answers, bound retries, and move to a flat plan once usage passes break-even.</description>
    </item>
    <item>
      <title>How to test an LLM for sensitive-data safety: BroadNet's method</title>
      <link>https://broadrouter.com/blog/llm-sensitive-data-safety-evaluation-method/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/llm-sensitive-data-safety-evaluation-method/</guid>
      <pubDate>Sun, 15 Nov 2026 08:00:00 GMT</pubDate>
      <category>Privacy</category>
      <description>BroadNet's April 2026 preprint (DOI 10.5281/zenodo.19574049, also on SSRN) sets out a test for sensitive-data safety before deployment. Test four categories (credentials, PII, health and financial data) separately, score leaks and misses as different failures, use binary scoring, and run every test at least three times. Prompt rules and fine-tuning on negative examples made leaks worse in its tests.</description>
    </item>
    <item>
      <title>Live, on request, research: read a model catalogue clearly</title>
      <link>https://broadrouter.com/blog/understand-model-status/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/understand-model-status/</guid>
      <pubDate>Sun, 15 Nov 2026 08:00:00 GMT</pubDate>
      <category>Comparisons</category>
      <description>A model you can call has an exact API ID, a published price and a live status on the provider's rate card or price feed. On-request models need an enquiry and usually a private model ID. Research entries record what the provider tested and are not callable. Read the feed and API reference, then confirm with one request.</description>
    </item>
    <item>
      <title>LLM API failover: fall back to a second provider in your own code</title>
      <link>https://broadrouter.com/blog/llm-api-failover-fallback/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/llm-api-failover-fallback/</guid>
      <pubDate>Sun, 15 Nov 2026 08:00:00 GMT</pubDate>
      <category>Engineering</category>
      <description>Keep an ordered list of OpenAI-compatible providers, each with its own client and model ID. Fail over on connection errors, timeouts, 5xx and rate-limit 429s once the SDK's own retries are spent. Never fail over on 400, 401, 402, 403, 413 or 422. Add a circuit breaker, and check that the backup matches the primary on features and data terms.</description>
    </item>
    <item>
      <title>LLM API price cuts 2023–2026: the biggest drops, with sources</title>
      <link>https://broadrouter.com/blog/llm-api-price-cuts-2023-2026/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/llm-api-price-cuts-2023-2026/</guid>
      <pubDate>Sun, 15 Nov 2026 08:00:00 GMT</pubDate>
      <category>Pricing</category>
      <description>The largest list-price cuts in our archive of 11 providers' pricing pages were 80% on OpenAI o3 (mid-2025) and Mistral Codestral (late 2024), 75% on Gemini 1.5 Flash and 81% on Mistral Small 3 at Together AI. Makers' cuts were large (median 58%); hosts' were small (median 14%). We also found 52 rises, including Claude 3.5 Haiku.</description>
    </item>
    <item>
      <title>LLM API rate limits by provider: RPM, TPM and concurrency compared</title>
      <link>https://broadrouter.com/blog/llm-api-rate-limits-by-provider/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/llm-api-rate-limits-by-provider/</guid>
      <pubDate>Sun, 15 Nov 2026 08:00:00 GMT</pubDate>
      <category>Engineering</category>
      <description>LLM API rate limits differ in kind, not just size. OpenAI, Anthropic and Gemini cap requests and tokens per minute by usage tier. DeepSeek and DeepInfra cap concurrent requests. Groq and Cerebras free tiers add daily caps. On first paid tiers, OpenAI and Anthropic allow thousands of requests a minute; BroadNet pay-as-you-go keys allow 30.</description>
    </item>
    <item>
      <title>LLM API SLAs compared: what Azure, Bedrock, Google and OpenAI promise</title>
      <link>https://broadrouter.com/blog/llm-api-sla-uptime-compared/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/llm-api-sla-uptime-compared/</guid>
      <pubDate>Sun, 15 Nov 2026 08:00:00 GMT</pubDate>
      <category>Comparisons</category>
      <description>Most LLM APIs promise no uptime on pay-as-you-go. Written SLAs come with cloud contracts or reserved capacity: Azure Foundry Models and Amazon Bedrock credit you below 99.9% monthly uptime, Google's Gemini online inference below 99.5%, and OpenAI's Scale Tier offers 99.9% to Enterprise buyers. Anthropic's commercial terms and BroadNet's self-serve terms have none. The remedy is a bill credit.</description>
    </item>
    <item>
      <title>LLM batch processing without a Batch API: a resumable runner</title>
      <link>https://broadrouter.com/blog/llm-batch-processing-without-batch-api/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/llm-batch-processing-without-batch-api/</guid>
      <pubDate>Sun, 15 Nov 2026 08:00:00 GMT</pubDate>
      <category>Engineering</category>
      <description>Without a Batch API, run the job yourself: read items from a JSONL file, cap requests in flight at the key's limit, pace request starts under the per-minute limit, retry 429 and 5xx with capped backoff, and append each result with its ID. On restart, skip finished IDs. The rate limit, not the model, usually sets the finish time.</description>
    </item>
    <item>
      <title>Make human review part of your document AI workflow</title>
      <link>https://broadrouter.com/blog/document-review-queues/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/document-review-queues/</guid>
      <pubDate>Sun, 15 Nov 2026 08:00:00 GMT</pubDate>
      <category>Engineering</category>
      <description>Treat every extracted document as unapproved until rules pass. Send it to a person when a required field is missing, sums disagree, the date or vendor looks wrong, or the model is unsure. Also review a small random share of clean documents, log every correction in your own system, and track the review rate and audited error rate.</description>
    </item>
    <item>
      <title>Migrating off the OpenAI API: a code walkthrough in Python, Node, cURL</title>
      <link>https://broadrouter.com/blog/migrate-from-openai-api-code-walkthrough/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/migrate-from-openai-api-code-walkthrough/</guid>
      <pubDate>Sun, 15 Nov 2026 08:00:00 GMT</pubDate>
      <category>Engineering</category>
      <description>Change the client's base URL, API key and model ID. That is enough for Chat Completions calls. Code on the Responses API must be rewritten to Chat Completions: input becomes messages, output_text becomes choices[0].message.content, and tools lose a level of nesting. Keep embeddings on a separate client if the new provider has none, then run a smoke test before switching traffic.</description>
    </item>
    <item>
      <title>Models BroadNet evaluated and ruled out, and why</title>
      <link>https://broadrouter.com/blog/models-broadnet-ruled-out/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/models-broadnet-ruled-out/</guid>
      <pubDate>Sun, 15 Nov 2026 08:00:00 GMT</pubDate>
      <category>Comparisons</category>
      <description>BroadNet's model catalogue marks about 30 models and methods as ruled out after testing in its lab. The reasons mostly fall into five groups: a model lost to a stronger one for the same job, a credential-safety gate, a licence that blocks API serving, a quantisation format that lost a head-to-head, and Arabic regressions. All results are BroadNet's own measurements from September 2026.</description>
    </item>
    <item>
      <title>Open-weight vs closed model API cost: what the price data shows</title>
      <link>https://broadrouter.com/blog/open-weight-vs-closed-model-api-cost/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/open-weight-vs-closed-model-api-cost/</guid>
      <pubDate>Sun, 15 Nov 2026 08:00:00 GMT</pubDate>
      <category>Pricing</category>
      <description>Usually, but not by model name alone. On prices checked on 15 November 2026, GLM 5.3 at its maker costs less per token than GPT-6.1 Sol or Claude Sonnet 5.5, yet small closed models such as GPT-6 Luna cost about the same as hosted open models. The host also matters: one open model can cost several times more at one host than another.</description>
    </item>
    <item>
      <title>OpenAI vs Anthropic vs Google API price history, 2023–2026</title>
      <link>https://broadrouter.com/blog/openai-anthropic-google-api-price-history/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/openai-anthropic-google-api-price-history/</guid>
      <pubDate>Sun, 15 Nov 2026 08:00:00 GMT</pubDate>
      <category>Comparisons</category>
      <description>Top-tier prices fell; small tiers at Anthropic and Google rose. In archived pricing pages, GPT-4 was USD 30 input and USD 60 output per 1M tokens (April 2023) and GPT-6.1 Sol USD 2 and USD 10 (October 2026). Claude Opus fell from USD 15 / 75 to USD 4 / 20. But Claude Haiku 4.5 costs four times Haiku 3.</description>
    </item>
    <item>
      <title>OpenAI-compatible API base URLs: 14 providers, checked</title>
      <link>https://broadrouter.com/blog/openai-compatible-api-base-urls/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/openai-compatible-api-base-urls/</guid>
      <pubDate>Sun, 15 Nov 2026 08:00:00 GMT</pubDate>
      <category>Engineering</category>
      <description>An OpenAI-compatible API lets you keep the OpenAI SDK and change two things: base_url and the API key. The base URL differs by provider, for example https://api.groq.com/openai/v1 for Groq, https://openrouter.ai/api/v1 for OpenRouter and https://inference.broadnet.ai/v1 for BroadNet Inference. Paths, model IDs and supported parameters still differ, so test one request before you switch.</description>
    </item>
    <item>
      <title>OpenClaw and Hermes Agent on a cheap model: setup and token math</title>
      <link>https://broadrouter.com/blog/openclaw-hermes-agent-cheap-model/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/openclaw-hermes-agent-cheap-model/</guid>
      <pubDate>Sun, 15 Nov 2026 08:00:00 GMT</pubDate>
      <category>Agents</category>
      <description>Both agents accept a custom OpenAI-compatible endpoint: OpenClaw under models.providers with api "openai-completions", Hermes Agent with provider: custom and a base_url. Cost is driven by context size times how often the agent runs times the model price. Shrink the context, run scheduled checks less often and on a cheaper model, and test tool calling before you commit.</description>
    </item>
    <item>
      <title>Prompt injection in coding agents: defences that actually hold</title>
      <link>https://broadrouter.com/blog/prompt-injection-coding-agents/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/prompt-injection-coding-agents/</guid>
      <pubDate>Sun, 15 Nov 2026 08:00:00 GMT</pubDate>
      <category>Agents</category>
      <description>Assume a coding agent will sometimes follow instructions hidden in a README, issue, web page or tool result; OWASP says fool-proof prevention may not exist. So limit what a hijacked agent can do: sandbox it with no network by default, keep secrets out, allow-list commands, require approval to push or install, and review every change.</description>
    </item>
    <item>
      <title>RAG chunk size and overlap: what they do to retrieval and cost</title>
      <link>https://broadrouter.com/blog/rag-chunk-size-overlap/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/rag-chunk-size-overlap/</guid>
      <pubDate>Sun, 15 Nov 2026 08:00:00 GMT</pubDate>
      <category>Engineering</category>
      <description>Start with chunks of about 256 to 512 tokens, 0 to 15% overlap, and top-k tuned on your own queries. Bigger chunks raise the tokens every query sends to the LLM. Overlap multiplies the tokens you embed and store by size ÷ (size − overlap), so 50% overlap doubles the index. Published tests found small or no-overlap chunks often recall as well or better.</description>
    </item>
    <item>
      <title>Record audio in the browser and send it to a speech-to-text API</title>
      <link>https://broadrouter.com/blog/record-browser-audio-transcription-api/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/record-browser-audio-transcription-api/</guid>
      <pubDate>Sun, 15 Nov 2026 08:00:00 GMT</pubDate>
      <category>Engineering</category>
      <description>Record with getUserMedia and MediaRecorder, picking a format with isTypeSupported, and drop empty chunks. If the API accepts only WAV or MP3, decode the recording in the browser and encode 16 kHz mono WAV, or upload it and convert with ffmpeg on your server. Either way, send it through your own backend so the API key never reaches the browser.</description>
    </item>
    <item>
      <title>Reduce prompt tokens: measured cuts for system prompts, tools, data</title>
      <link>https://broadrouter.com/blog/reduce-prompt-tokens/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/reduce-prompt-tokens/</guid>
      <pubDate>Sun, 15 Nov 2026 08:00:00 GMT</pubDate>
      <category>Engineering</category>
      <description>Measure first: log prompt_tokens per call, find the biggest part, and diff with the model's tokenizer. Across five tokenizers, trimming a wordy system prompt saved 68 to 71% and CSV used 31 to 41% fewer tokens than minified JSON. On three models, a trimmed prompt with 3 short tools instead of 10 verbose ones cut the full request by about 65 to 71%.</description>
    </item>
    <item>
      <title>Same model, different host: how much LLM API prices vary</title>
      <link>https://broadrouter.com/blog/same-llm-different-host-prices/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/same-llm-different-host-prices/</guid>
      <pubDate>Sun, 15 Nov 2026 08:00:00 GMT</pubDate>
      <category>Comparisons</category>
      <description>A lot. Across 119 models that two or more independent hosts sell, the dearest list price is a median 1.76 times the cheapest (blended, 3:1), from 41 providers' pages checked 14 and 15 November 2026. 50 models have a gap of 2 times or more. Yet most hosts copy the maker: 57% of host prices equal the maker's own price exactly.</description>
    </item>
    <item>
      <title>Saudi and UAE PDPL and LLM APIs: a cross-border checklist</title>
      <link>https://broadrouter.com/blog/saudi-uae-pdpl-llm-api-checklist/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/saudi-uae-pdpl-llm-api-checklist/</guid>
      <pubDate>Sun, 15 Nov 2026 08:00:00 GMT</pubDate>
      <category>Privacy</category>
      <description>Under Saudi Arabia's PDPL, sending personal data in a prompt to an LLM API hosted abroad is a transfer outside the Kingdom. Article 29 and SDAIA's transfer regulation set the conditions: an adequate destination or safeguards, the minimum data needed, and a risk assessment in some cases. The UAE law also regulates cross-border transfers. Minimise first, then check where the provider processes. Not legal advice.</description>
    </item>
    <item>
      <title>Setting max_tokens for reasoning models: a budgeting method</title>
      <link>https://broadrouter.com/blog/reasoning-model-max-tokens-budget/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/reasoning-model-max-tokens-budget/</guid>
      <pubDate>Sun, 15 Nov 2026 08:00:00 GMT</pubDate>
      <category>Engineering</category>
      <description>On a reasoning model, max_tokens caps thinking and answer together, so a tight limit can end the reply before any visible text. Measure total completion tokens per task type, set max_tokens to about the 95th percentile plus 20%, keep prompt plus max_tokens inside the context window, and retry cut-off replies with a doubled budget up to a hard cap.</description>
    </item>
    <item>
      <title>Text, speech or vision? Start with the task</title>
      <link>https://broadrouter.com/blog/text-speech-vision/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/text-speech-vision/</guid>
      <pubDate>Sun, 15 Nov 2026 08:00:00 GMT</pubDate>
      <category>Engineering</category>
      <description>Start from your input and output. Text in, text out is chat. A photo or scan in, text out is chat with an image, on a model that accepts images. Recorded speech in, text out is transcription; speech to an answer is transcription then chat. BroadNet does not offer text-to-speech, image generation or embeddings.</description>
    </item>
    <item>
      <title>The Safety Map: how 39 LLMs handle sensitive data</title>
      <link>https://broadrouter.com/blog/llm-safety-map-39-models/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/llm-safety-map-39-models/</guid>
      <pubDate>Sun, 15 Nov 2026 08:00:00 GMT</pubDate>
      <category>Privacy</category>
      <description>The Safety Map is a BroadNet preprint from April 2026 (DOI 10.5281/zenodo.19688433). It tested 39 models from 14 labs on 7 probes across credentials, PII, health data, account numbers and DLP scanning, in about 1,300 multi-run evaluations. Safety did not transfer between categories or output channels: 20 of 73 runs that filled tool-call arguments leaked through them.</description>
    </item>
    <item>
      <title>Transcribe long audio over 25 MB: split, send and stitch</title>
      <link>https://broadrouter.com/blog/transcribe-long-audio-files-api/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/transcribe-long-audio-files-api/</guid>
      <pubDate>Sun, 15 Nov 2026 08:00:00 GMT</pubDate>
      <category>Engineering</category>
      <description>Most transcription APIs reject one file over about 25 MB. Convert the recording to 16 kHz mono first: that alone fits about 13 minutes of WAV or 54 minutes of 64 kbit/s MP3 into BroadNet's 26,214,400-byte limit. For longer audio, split at silences with ffmpeg, transcribe the chunks in parallel and join the text in order.</description>
    </item>
    <item>
      <title>Translate app strings to Arabic with an LLM: placeholders and plurals</title>
      <link>https://broadrouter.com/blog/translate-app-strings-to-arabic-llm/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/translate-app-strings-to-arabic-llm/</guid>
      <pubDate>Sun, 15 Nov 2026 08:00:00 GMT</pubDate>
      <category>Arabic AI</category>
      <description>Send the strings as JSON with every placeholder swapped for a marker, ask for all six Arabic plural forms (zero, one, two, few, many, other), and build the ICU plural in code, not in the model. Validate in CI: same keys, same arguments and tags, all six forms present. Wrap inserted names in Unicode isolates so right-to-left text displays correctly.</description>
    </item>
    <item>
      <title>Translate Arabic audio to English text with an API: a tested pipeline</title>
      <link>https://broadrouter.com/blog/translate-arabic-audio-to-english/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/translate-arabic-audio-to-english/</guid>
      <pubDate>Sun, 15 Nov 2026 08:00:00 GMT</pubDate>
      <category>Arabic AI</category>
      <description>Run two steps. Transcribe the Arabic audio with a speech-to-text API, setting language to ar, then translate the transcript with an LLM that is told to keep names and numbers. Split long files first, check numbers in code, and score a small sample with chrF to see how much the transcription step costs you in quality.</description>
    </item>
    <item>
      <title>Uno in vLLM: what BroadNet's serving study measured</title>
      <link>https://broadrouter.com/blog/uno-in-vllm-serving-study/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/uno-in-vllm-serving-study/</guid>
      <pubDate>Sun, 15 Nov 2026 08:00:00 GMT</pubDate>
      <category>Engineering</category>
      <description>Uno in vLLM is a September 2026 preprint by BroadNet (DOI 10.5281/zenodo.22652610). It adds Uno, an adapter that lets a model draft its own tokens, to vLLM. On an H100 with Qwen3-8B, BroadNet measured 2.58 times plain-vLLM throughput at one stream, 2.20 times at eight and 1.74 times at 32, within 3% of the reference implementation.</description>
    </item>
    <item>
      <title>Uno on Gemma 4 26B: BroadNet's mixture-of-experts study</title>
      <link>https://broadrouter.com/blog/uno-mixture-of-experts-gemma-4-study/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/uno-mixture-of-experts-gemma-4-study/</guid>
      <pubDate>Sun, 15 Nov 2026 08:00:00 GMT</pubDate>
      <category>Engineering</category>
      <description>BroadNet's September 2026 preprint (DOI 10.5281/zenodo.22820511) trains the first Uno adapter for a mixture-of-experts model, Gemma 4 26B A4B, served at 4-bit. The adapter decodes 1.166 times faster than plain decoding on a 72-prompt production workload, behind two separate drafters at 1.474 and 1.553, but leads the production drafter at about 10,500 tokens of context.</description>
    </item>
    <item>
      <title>What one AI coding agent run costs: steps, context growth, caching</title>
      <link>https://broadrouter.com/blog/ai-coding-agent-run-cost/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/ai-coding-agent-run-cost/</guid>
      <pubDate>Sun, 15 Nov 2026 08:00:00 GMT</pubDate>
      <category>Agents</category>
      <description>An agent resends its whole context on every step, so input grows with the square of the step count: 60 steps cost about 3.5 times the input of 30. Caching decides most of the bill; with append-only history, over 90% of input can bill at the cached rate. Multiply cost per run by monthly runs and compare with a flat session.</description>
    </item>
    <item>
      <title>10 ways to cut coding agent LLM costs, with estimated savings</title>
      <link>https://broadrouter.com/blog/cut-coding-agent-llm-costs/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/cut-coding-agent-llm-costs/</guid>
      <pubDate>Sat, 14 Nov 2026 08:00:00 GMT</pubDate>
      <category>Pricing</category>
      <description>Most spend is input resent every turn. Keep the prompt prefix stable so it caches, name files, trim tool output, lower reasoning effort, cap loops and retries, and route easy steps to cheaper models. On DeepSeek V4.1 Flash, cached input costs about 3% of new input. Past about 917 million input tokens a month, a $60 flat plan is cheaper.</description>
    </item>
    <item>
      <title>A weekly LLM plan for short projects: Orbit vs hour vs month</title>
      <link>https://broadrouter.com/blog/weekly-llm-plan-short-projects/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/weekly-llm-plan-short-projects/</guid>
      <pubDate>Sat, 14 Nov 2026 08:00:00 GMT</pubDate>
      <category>Pricing</category>
      <description>Orbit is one week of unlimited DeepSeek V4.1 Flash for $15 per session, paid from your BroadNet balance, with no token meter and one request at a time. It fits a sprint, hackathon or migration. For an afternoon, Meteor at $1 an hour is cheaper; for continuous use, Full Moon at $60 beats back-to-back weeks ($64.29 over 30 days).</description>
    </item>
    <item>
      <title>Agent token budget: where an agent loop's tokens go</title>
      <link>https://broadrouter.com/blog/agent-token-budget/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/agent-token-budget/</guid>
      <pubDate>Sat, 14 Nov 2026 08:00:00 GMT</pubDate>
      <category>Agents</category>
      <description>An agent loop resends its whole context on every call: system prompt, tool definitions, conversation history, file contents and tool output, plus the new step. Input usually outweighs output many times over. Log prompt_tokens, cached_tokens and completion_tokens per call, find the biggest repeating block, then trim tools, cap tool output and restart long sessions.</description>
    </item>
    <item>
      <title>Arabic ASR word error rate (WER) explained</title>
      <link>https://broadrouter.com/blog/arabic-asr-wer-explained/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/arabic-asr-wer-explained/</guid>
      <pubDate>Sat, 14 Nov 2026 08:00:00 GMT</pubDate>
      <category>Arabic AI</category>
      <description>Word error rate (WER) is the number of substituted, deleted and inserted words divided by the words in the reference transcript. Lower is better. Arabic WER is hard to compare because diacritics, alef and ta marbuta spelling and dialects inflate it. BroadNet reports 3.8 WER on FLEURS-ar and 23.4 on dialect speech for BroadNet Transcribe Arabic.</description>
    </item>
    <item>
      <title>Arabic dialect speech recognition: Gulf, Egyptian, Levantine, Maghrebi</title>
      <link>https://broadrouter.com/blog/arabic-dialect-speech-recognition/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/arabic-dialect-speech-recognition/</guid>
      <pubDate>Sat, 14 Nov 2026 08:00:00 GMT</pubDate>
      <category>Arabic AI</category>
      <description>Arabic dialects differ from Modern Standard Arabic in sounds, words and spelling, and speakers mix in English or French, so ASR models trained mostly on formal Arabic score far worse on them. To test a provider, build a test set per dialect from your real audio, fix a spelling convention, and score WER, CER and entity accuracy per dialect.</description>
    </item>
    <item>
      <title>Arabic OCR vs vision LLMs: reading Arabic documents and invoices</title>
      <link>https://broadrouter.com/blog/arabic-ocr-vision-llm/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/arabic-ocr-vision-llm/</guid>
      <pubDate>Sat, 14 Nov 2026 08:00:00 GMT</pubDate>
      <category>Arabic AI</category>
      <description>Classic OCR turns Arabic page images into raw text; a vision LLM reads the image and returns the fields you ask for, such as vendor, date and total, as JSON. For invoices and receipts a vision LLM usually saves the parsing step. On BroadNet Inference, Qwen3.8 27B, DeepSeek V4.1 Flash and BroadNet Gemma 4 VL 26B accept images.</description>
    </item>
    <item>
      <title>Arabic prompt engineering: register, RTL, numerals and mixed text</title>
      <link>https://broadrouter.com/blog/arabic-prompt-engineering/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/arabic-prompt-engineering/</guid>
      <pubDate>Sat, 14 Nov 2026 08:00:00 GMT</pubDate>
      <category>Arabic AI</category>
      <description>Good Arabic output comes from explicit instructions: name the register (MSA or a specific dialect), the audience, the numeral style (Western 0–9 or Eastern Arabic ٠–٩), the formatting allowed, and how to treat English terms. Render replies with dir="auto" or isolation marks so mixed Arabic and English displays in the right order.</description>
    </item>
    <item>
      <title>Arabic RAG: normalisation, chunking and retrieval that works</title>
      <link>https://broadrouter.com/blog/arabic-rag-search/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/arabic-rag-search/</guid>
      <pubDate>Sat, 14 Nov 2026 08:00:00 GMT</pubDate>
      <category>Arabic AI</category>
      <description>Arabic RAG works when you normalise text before indexing (unify alef forms, ya and alef maqsura, ta marbuta, strip diacritics and tatweel), chunk by paragraph, combine a multilingual embedding model with keyword search, and let a long-context LLM answer from the retrieved passages with citations. Apply the same normalisation to queries.</description>
    </item>
    <item>
      <title>Arabic speech-to-text API: how to choose one in 2026</title>
      <link>https://broadrouter.com/blog/arabic-speech-to-text-api-buyers-guide/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/arabic-speech-to-text-api-buyers-guide/</guid>
      <pubDate>Sat, 14 Nov 2026 08:00:00 GMT</pubDate>
      <category>Arabic AI</category>
      <description>Pick an Arabic speech-to-text API by testing it on your own recordings: measure word error rate on your dialects, then compare price per audio hour, file limits and data retention. BroadNet Transcribe Arabic costs $0.25 per audio hour ($0.0042 a minute), and the Open Universal Arabic ASR Leaderboard listed it #4 of 39 on 15 November 2026.</description>
    </item>
    <item>
      <title>Arabic WhatsApp chatbot architecture: voice notes to replies</title>
      <link>https://broadrouter.com/blog/arabic-whatsapp-chatbot-architecture/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/arabic-whatsapp-chatbot-architecture/</guid>
      <pubDate>Sat, 14 Nov 2026 08:00:00 GMT</pubDate>
      <category>Arabic AI</category>
      <description>An Arabic WhatsApp assistant has five parts: a webhook that receives messages from the WhatsApp Business Platform, a queue, speech to text for voice notes, an LLM that drafts the reply, and a sender that answers through the API. Converting a voice note and transcribing it with BroadNet Transcribe Arabic costs $0.0042 per audio minute.</description>
    </item>
    <item>
      <title>Arabic–English translation with LLMs: prompts, glossaries, evaluation</title>
      <link>https://broadrouter.com/blog/arabic-english-translation-llm/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/arabic-english-translation-llm/</guid>
      <pubDate>Sat, 14 Nov 2026 08:00:00 GMT</pubDate>
      <category>Arabic AI</category>
      <description>To translate Arabic and English with an LLM, give the model the audience, register and a glossary of fixed terms, ask it to keep numbers and names unchanged, and translate in paragraph-sized chunks. Evaluate with chrF or COMET plus native review. In BroadNet's tests, Qwen3.8 27B translated accurately between Arabic and English.</description>
    </item>
    <item>
      <title>Arabizi and LLMs: handling Franco-Arabic with numbers like 3 and 7</title>
      <link>https://broadrouter.com/blog/arabizi-llm-handling/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/arabizi-llm-handling/</guid>
      <pubDate>Sat, 14 Nov 2026 08:00:00 GMT</pubDate>
      <category>Arabic AI</category>
      <description>Arabizi, also called Franco-Arabic, is Arabic written in Latin letters with digits for sounds Latin lacks: 3 for ع, 7 for ح, 2 for ء, 5 for خ. LLMs usually understand it but convert it inconsistently. Detect it, ask the model to transliterate into Arabic script with the dialect named, then process the Arabic text.</description>
    </item>
    <item>
      <title>Best LLM for Aider: how to choose a model and connect it</title>
      <link>https://broadrouter.com/blog/best-llm-for-aider/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/best-llm-for-aider/</guid>
      <pubDate>Sat, 14 Nov 2026 08:00:00 GMT</pubDate>
      <category>Agents</category>
      <description>The best model for Aider follows its edit format reliably, holds your repository map in context and is cheap enough to run all day. Aider connects to any OpenAI-compatible API: set OPENAI_API_BASE to https://inference.broadnet.ai/v1, set OPENAI_API_KEY, and start aider --model openai/broadnet/deepseek-v4.1-flash. On a Full Moon plan that model has no token meter.</description>
    </item>
    <item>
      <title>Best open model for coding agents: pick by tools, context and cost</title>
      <link>https://broadrouter.com/blog/best-open-model-for-coding-agents/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/best-open-model-for-coding-agents/</guid>
      <pubDate>Sat, 14 Nov 2026 08:00:00 GMT</pubDate>
      <category>Agents</category>
      <description>For a coding agent that depends on native tool calls (opencode, Continue Agent mode, Zoo Code), GLM 5.3 is the safest pick on BroadNet: tools listed, 1M context. For Aider or an all-day agent, DeepSeek V4.1 Flash on a flat plan costs least. Qwen3.8 27B fits screenshot work. Gemma 4 VL 26B's 32K context is too small.</description>
    </item>
    <item>
      <title>Cached input tokens: how to cut an LLM API bill</title>
      <link>https://broadrouter.com/blog/cached-input-tokens-explained/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/cached-input-tokens-explained/</guid>
      <pubDate>Sat, 14 Nov 2026 08:00:00 GMT</pubDate>
      <category>Pricing</category>
      <description>Cached input tokens are prompt tokens the provider has already processed and can reuse, so they are billed at a lower rate. On BroadNet Inference, DeepSeek V4.1 Flash bills cached input at $0.0055 per 1M tokens against $0.175 for new input. Keep the stable part of every prompt first and identical between calls to get the discount.</description>
    </item>
    <item>
      <title>Cheapest vision LLM API: cost per image and per page, worked out</title>
      <link>https://broadrouter.com/blog/cheapest-vision-llm-api/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/cheapest-vision-llm-api/</guid>
      <pubDate>Sat, 14 Nov 2026 08:00:00 GMT</pubDate>
      <category>Pricing</category>
      <description>A vision LLM bills an image as input tokens, so 1,000 images of about 1,000 tokens each cost the input price per 1M tokens. On BroadNet the cheapest is Gemma 4 VL 26B at $0.10 input and $0.22 output per 1M. DeepInfra lists Google's Gemma 4 26B A4B with lower input and higher output prices.</description>
    </item>
    <item>
      <title>Choosing an LLM API for Arabic: what to test</title>
      <link>https://broadrouter.com/blog/choosing-an-llm-for-arabic/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/choosing-an-llm-for-arabic/</guid>
      <pubDate>Sat, 14 Nov 2026 08:00:00 GMT</pubDate>
      <category>Arabic AI</category>
      <description>Choose an LLM for Arabic by testing it on your own tasks: understanding Gulf, Egyptian and Levantine dialects, writing fluent Modern Standard Arabic, translating Arabic and English, and reading Arabic documents. In BroadNet's tests, Qwen3.8 27B wrote fluent Arabic and understood Gulf and Egyptian dialects. It costs $0.175 per 1M input tokens.</description>
    </item>
    <item>
      <title>Cline vs Roo Code vs Aider for teams on open models (2026)</title>
      <link>https://broadrouter.com/blog/cline-vs-roo-code-vs-aider/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/cline-vs-roo-code-vs-aider/</guid>
      <pubDate>Sat, 14 Nov 2026 08:00:00 GMT</pubDate>
      <category>Comparisons</category>
      <description>Pick Cline for a maintained VS Code agent with approvals, checkpoints, MCP and Plan/Act models. Pick Aider for a git-first terminal workflow that edits through text formats and so runs on models without tool calling. Roo Code was archived on May 15, 2026; existing users can move to Zoo Code, its community continuation, or to Cline.</description>
    </item>
    <item>
      <title>Context management for long-running agents with a 1M window</title>
      <link>https://broadrouter.com/blog/long-running-agent-context-management/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/long-running-agent-context-management/</guid>
      <pubDate>Sat, 14 Nov 2026 08:00:00 GMT</pubDate>
      <category>Agents</category>
      <description>Treat the context as an append-only log: a fixed system prompt and tool list first, then history that only grows, so each turn reuses the cached prefix. With a 1,048,576-token window, as on DeepSeek V4.1 Flash, compact rarely and in large steps, trim bulky tool output first, and keep durable facts in a short notes file the agent rereads.</description>
    </item>
    <item>
      <title>DeepSeek API alternatives: 5 ways to call DeepSeek models</title>
      <link>https://broadrouter.com/blog/deepseek-api-alternatives/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/deepseek-api-alternatives/</guid>
      <pubDate>Sat, 14 Nov 2026 08:00:00 GMT</pubDate>
      <category>Comparisons</category>
      <description>You can call DeepSeek models through DeepSeek's own API, an aggregator that routes to several hosts, a cloud provider's model catalogue, your own GPUs, or an independent provider such as BroadNet Inference. BroadNet sells DeepSeek V4.1 Flash, served by a US inference partner with zero data retention, at $0.175 per 1M input tokens or unlimited for $60 a month per session.</description>
    </item>
    <item>
      <title>DeepSeek V4.1 Flash API pricing: per token, cached and plans</title>
      <link>https://broadrouter.com/blog/deepseek-v4-1-flash-api-pricing/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/deepseek-v4-1-flash-api-pricing/</guid>
      <pubDate>Sat, 14 Nov 2026 08:00:00 GMT</pubDate>
      <category>Pricing</category>
      <description>On BroadNet Inference, DeepSeek V4.1 Flash costs $0.175 per 1M input tokens, $0.0055 per 1M cached input tokens and $0.52 per 1M output tokens, reasoning included. Or pay a flat price per session: $60 a month, $15 a week or $1 an hour, with no token meter, one request at a time.</description>
    </item>
    <item>
      <title>Extract invoice, receipt and form data to JSON with a vision LLM</title>
      <link>https://broadrouter.com/blog/extract-data-from-documents-with-vision-llm/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/extract-data-from-documents-with-vision-llm/</guid>
      <pubDate>Sat, 14 Nov 2026 08:00:00 GMT</pubDate>
      <category>Engineering</category>
      <description>Render each page to an image, send it to a vision model with a prompt that names every field and its format, and ask for JSON only. Then validate in code: parse it, check its schema, check that line items add up to the total, and retry once with the error. Route repeat failures to a person.</description>
    </item>
    <item>
      <title>Flat-rate vs pay-per-token LLM API: the break-even math</title>
      <link>https://broadrouter.com/blog/flat-rate-vs-pay-per-token-llm-api/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/flat-rate-vs-pay-per-token-llm-api/</guid>
      <pubDate>Sat, 14 Nov 2026 08:00:00 GMT</pubDate>
      <category>Pricing</category>
      <description>For DeepSeek V4.1 Flash on BroadNet Inference, a flat Full Moon plan ($60 a month) beats pay-per-token at roughly 917 million input and 46 million output tokens a month, assuming 80% cache hits. A full-time coding agent usually crosses that line. A side project or chat assistant usually does not.</description>
    </item>
    <item>
      <title>Free LLM API credit: how far a starting credit really goes</title>
      <link>https://broadrouter.com/blog/free-llm-api-credit/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/free-llm-api-credit/</guid>
      <pubDate>Sat, 14 Nov 2026 08:00:00 GMT</pubDate>
      <category>Pricing</category>
      <description>New BroadNet Inference accounts get $5 of free pay-as-you-go credit. It works on BroadNet Gemma 4 VL 26B ($0.10 per 1M input tokens), the transcription models and hourly keys. That is tens of millions of Gemma tokens, or about 1,190 minutes of Arabic transcription. DeepSeek V4.1 Flash, GLM 5.3 and Qwen3.8 27B open after your first top-up.</description>
    </item>
    <item>
      <title>GDPR and LLM APIs: 15 questions to ask your provider</title>
      <link>https://broadrouter.com/blog/gdpr-llm-api-checklist/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/gdpr-llm-api-checklist/</guid>
      <pubDate>Sat, 14 Nov 2026 08:00:00 GMT</pubDate>
      <category>Privacy</category>
      <description>Before you send personal data to an LLM API, get a signed data processing agreement (GDPR Article 28), the subprocessor list, written retention and training terms, the processing location and the legal basis for any transfer outside the EEA (Articles 44 to 49). Then minimise what you send. This is a checklist, not legal advice.</description>
    </item>
    <item>
      <title>GLM 5.3 vs DeepSeek V4.1 Flash: which one to use, and when</title>
      <link>https://broadrouter.com/blog/glm-5-3-vs-deepseek-v4-1-flash/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/glm-5-3-vs-deepseek-v4-1-flash/</guid>
      <pubDate>Sat, 14 Nov 2026 08:00:00 GMT</pubDate>
      <category>Comparisons</category>
      <description>Pick DeepSeek V4.1 Flash by default: it costs $0.175 input and $0.52 output per 1M tokens, reads images, and runs unlimited on flat plans. Pick GLM 5.3 ($0.78 and $3.46) when you need tool calling and structured output that the provider lists for the model. Both have a 1,048,576-token context and reason before answering.</description>
    </item>
    <item>
      <title>Gulf, Egyptian and Levantine Arabic: how LLMs handle dialects</title>
      <link>https://broadrouter.com/blog/arabic-llm-gulf-egyptian-dialects/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/arabic-llm-gulf-egyptian-dialects/</guid>
      <pubDate>Sat, 14 Nov 2026 08:00:00 GMT</pubDate>
      <category>Arabic AI</category>
      <description>LLMs handle Arabic dialects unevenly: they usually understand Egyptian and Gulf Arabic better than they write them, and quality drops for rarer dialects. Test understanding and writing separately on your own messages. In BroadNet's own tests, Qwen3.8 27B understood Gulf and Egyptian dialects and wrote fluent Arabic. It costs $0.175 per 1M input tokens.</description>
    </item>
    <item>
      <title>Hourly LLM API key: a flat price per hour, explained</title>
      <link>https://broadrouter.com/blog/hourly-llm-api-key-explained/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/hourly-llm-api-key-explained/</guid>
      <pubDate>Sat, 14 Nov 2026 08:00:00 GMT</pubDate>
      <category>Pricing</category>
      <description>An hourly key is a prepaid API key that works for one hour at a flat price. On BroadNet Inference it costs $0.50 for one hour of Gemma 4 VL 26B: one stream, 120 requests a minute and 300,000 tokens a minute, with no token meter. It suits demos, workshops and short batch jobs with a fixed cost.</description>
    </item>
    <item>
      <title>How inference providers price open models: 6 compared</title>
      <link>https://broadrouter.com/blog/llm-inference-provider-pricing-compared/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/llm-inference-provider-pricing-compared/</guid>
      <pubDate>Sat, 14 Nov 2026 08:00:00 GMT</pubDate>
      <category>Comparisons</category>
      <description>Most providers price open models per million tokens, with cheaper cached input and often 50% off batch jobs. Together AI, Fireworks and DeepInfra also rent GPUs by the hour. OpenRouter passes provider prices through and charges a fee on credit purchases. Groq covers a short model list. BroadNet adds flat monthly, weekly and hourly plans for DeepSeek V4.1 Flash.</description>
    </item>
    <item>
      <title>How to choose an LLM API provider: a 20-point checklist</title>
      <link>https://broadrouter.com/blog/choose-llm-api-provider-checklist/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/choose-llm-api-provider-checklist/</guid>
      <pubDate>Sat, 14 Nov 2026 08:00:00 GMT</pubDate>
      <category>Comparisons</category>
      <description>Choose an LLM API provider by scoring it on six areas: total cost for your real workload, rate and concurrency limits, OpenAI compatibility, data retention and training terms, the models and context sizes you need, and support and reliability. Then run a one-week test with your own traffic before committing. Price per token alone is a poor guide.</description>
    </item>
    <item>
      <title>How to count tokens before you send an LLM request</title>
      <link>https://broadrouter.com/blog/count-tokens-before-sending/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/count-tokens-before-sending/</guid>
      <pubDate>Sat, 14 Nov 2026 08:00:00 GMT</pubDate>
      <category>Engineering</category>
      <description>Count tokens with the tokenizer of the model you call when it is public; otherwise calibrate a characters-per-token ratio from the usage field of real requests, per language and content type. Add a 10 to 20 percent margin, then check that prompt tokens plus max_tokens fit the context window, leaving room for reasoning tokens on thinking models.</description>
    </item>
    <item>
      <title>How to estimate LLM API cost before you launch a feature</title>
      <link>https://broadrouter.com/blog/estimate-llm-api-cost-before-launch/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/estimate-llm-api-cost-before-launch/</guid>
      <pubDate>Sat, 14 Nov 2026 08:00:00 GMT</pubDate>
      <category>Pricing</category>
      <description>Estimate LLM API cost by modelling one user action, not one API call: count calls per action, measure input, cached and output tokens per call on real samples, then multiply by users and actions per month. Add reasoning and retries, and price low, expected and high scenarios. Our worked example on DeepSeek V4.1 Flash comes to $42.67 a month.</description>
    </item>
    <item>
      <title>How to evaluate an LLM before switching providers</title>
      <link>https://broadrouter.com/blog/evaluate-llm-before-switching/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/evaluate-llm-before-switching/</guid>
      <pubDate>Sat, 14 Nov 2026 08:00:00 GMT</pubDate>
      <category>Engineering</category>
      <description>Collect 50 to 200 real prompts with known good outcomes. Score both models with code checks first and an LLM judge only where code cannot decide. Swap answer order to cancel position bias, use a judge from a third model family, check it against your own labels, and fix the pass rule, including cost and latency, before you run it.</description>
    </item>
    <item>
      <title>How to handle LLM API 429 errors: backoff, jitter and limiters</title>
      <link>https://broadrouter.com/blog/handle-llm-api-429-rate-limits/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/handle-llm-api-429-rate-limits/</guid>
      <pubDate>Sat, 14 Nov 2026 08:00:00 GMT</pubDate>
      <category>Engineering</category>
      <description>A 429 means you hit a rate, concurrency or spend limit. Retry it with capped exponential backoff and random jitter, and honour Retry-After when sent. Then prevent most 429s: cap requests in flight with a semaphore at the key's limit and queue the rest. Retry 5xx the same way; fix other 4xx errors instead.</description>
    </item>
    <item>
      <title>How to run coding agents unattended overnight, safely</title>
      <link>https://broadrouter.com/blog/overnight-coding-agent-runs/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/overnight-coding-agent-runs/</guid>
      <pubDate>Sat, 14 Nov 2026 08:00:00 GMT</pubDate>
      <category>Agents</category>
      <description>Run the agent in a container or VM with only the repository mounted, on its own git branch, committing after every completed step. Give it a written task file, a step cap and a wall-clock deadline, then review the branch in the morning. A flat plan session, such as BroadNet's $60 Full Moon, fixes the worst-case bill.</description>
    </item>
    <item>
      <title>How to run parallel coding agents on a flat-rate LLM API</title>
      <link>https://broadrouter.com/blog/run-parallel-coding-agents-flat-rate/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/run-parallel-coding-agents-flat-rate/</guid>
      <pubDate>Sat, 14 Nov 2026 08:00:00 GMT</pubDate>
      <category>Agents</category>
      <description>Give each agent that works at the same time its own plan session. A session runs one request at a time with no token meter, so an orchestrator with three sub-agents needs four sessions. Cap requests in flight in your own code to match. One key can hold several sessions, and each session on Full Moon costs $60 a month.</description>
    </item>
    <item>
      <title>How to structure prompts so the prefix caches</title>
      <link>https://broadrouter.com/blog/prompt-structure-for-caching/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/prompt-structure-for-caching/</guid>
      <pubDate>Sat, 14 Nov 2026 08:00:00 GMT</pubDate>
      <category>Engineering</category>
      <description>Order every request from most stable to least stable: system prompt, tool definitions, reference documents, conversation history, then the new message. Make each layer byte-identical between calls: no timestamps or IDs at the top, tools sorted, JSON serialised deterministically, and history only ever appended. Then measure cached_tokens to confirm the prefix is actually reused.</description>
    </item>
    <item>
      <title>How to transcribe call center recordings: an Arabic-ready pipeline</title>
      <link>https://broadrouter.com/blog/transcribe-call-center-recordings/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/transcribe-call-center-recordings/</guid>
      <pubDate>Sat, 14 Nov 2026 08:00:00 GMT</pubDate>
      <category>Arabic AI</category>
      <description>To transcribe call center recordings, split stereo files into agent and customer channels, cut each channel into speech turns, transcribe every turn with the right language code, then merge the turns by start time. Redact card numbers and phone numbers from the transcript before an LLM summarises the call. Mono recordings need a separate diarization step.</description>
    </item>
    <item>
      <title>JSON mode vs json_schema: structured output that holds up</title>
      <link>https://broadrouter.com/blog/structured-output-json-schema/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/structured-output-json-schema/</guid>
      <pubDate>Sat, 14 Nov 2026 08:00:00 GMT</pubDate>
      <category>Engineering</category>
      <description>JSON mode (response_format json_object) only promises valid JSON. Structured output (response_format json_schema) also promises your fields and types, where the provider enforces it. Either way, check finish_reason, parse, and validate against a schema such as a Pydantic model. If validation fails, send the error back and ask again, with a hard limit on attempts.</description>
    </item>
    <item>
      <title>LLM API data retention compared: OpenAI, Anthropic, Gemini, Mistral</title>
      <link>https://broadrouter.com/blog/llm-api-data-retention-compared/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/llm-api-data-retention-compared/</guid>
      <pubDate>Sat, 14 Nov 2026 08:00:00 GMT</pubDate>
      <category>Privacy</category>
      <description>As of 14 November 2026, OpenAI keeps API abuse-monitoring logs for up to 30 days and Anthropic deletes API content within 30 days; neither trains on it by default. Mistral keeps it 30 rolling days for abuse monitoring. Google logs paid Gemini API prompts for "a limited period" without training on them. BroadNet Inference says it keeps no content and does not train.</description>
    </item>
    <item>
      <title>LLM API key security: storage, rotation and IP allow-lists</title>
      <link>https://broadrouter.com/blog/llm-api-key-security/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/llm-api-key-security/</guid>
      <pubDate>Sat, 14 Nov 2026 08:00:00 GMT</pubDate>
      <category>Engineering</category>
      <description>Keep LLM API keys in a secret manager, never in code or the browser. Give each app or agent its own key, rotate on a schedule, and lock keys to your servers with an IP allow-list. BroadNet Inference says its allow-lists take 32 entries, revocation stops new requests within five seconds, and keys last one year.</description>
    </item>
    <item>
      <title>LLM API pricing explained: input, cached, output and reasoning</title>
      <link>https://broadrouter.com/blog/llm-api-pricing-explained/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/llm-api-pricing-explained/</guid>
      <pubDate>Sat, 14 Nov 2026 08:00:00 GMT</pubDate>
      <category>Pricing</category>
      <description>LLM APIs bill per million tokens, with separate prices for input, cached input and output. Reasoning tokens are billed as output, images count as input tokens, and speech to text is billed per audio minute. On BroadNet Inference, DeepSeek V4.1 Flash costs $0.175 input, $0.0055 cached and $0.52 output per 1M tokens.</description>
    </item>
    <item>
      <title>LLM API spend control: budgets, key limits and loop protection</title>
      <link>https://broadrouter.com/blog/llm-api-spend-control/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/llm-api-spend-control/</guid>
      <pubDate>Sat, 14 Nov 2026 08:00:00 GMT</pubDate>
      <category>Pricing</category>
      <description>Control LLM API spend in layers: a prepaid balance as the account ceiling, one key per app locked to your server IPs, max_tokens and step caps on every request, and an in-code budget that stops a job before it overspends. Log tokens per key and alert on your own numbers. Flat plans and hourly keys cap the worst case.</description>
    </item>
    <item>
      <title>LLM API usage logging: cost per user and feature, PII redacted</title>
      <link>https://broadrouter.com/blog/llm-api-usage-logging/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/llm-api-usage-logging/</guid>
      <pubDate>Sat, 14 Nov 2026 08:00:00 GMT</pubDate>
      <category>Engineering</category>
      <description>Wrap every LLM call in one function that records model, feature, a hashed user ID, prompt, cached, reasoning and output tokens, cost, latency, status and the provider request ID. Compute cost from a price table, not the invoice. Log metadata by default; if you must keep prompt text, redact it before writing and set a short retention.</description>
    </item>
    <item>
      <title>LLM cost per user: unit economics for a SaaS AI feature</title>
      <link>https://broadrouter.com/blog/llm-cost-per-user-saas/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/llm-cost-per-user-saas/</guid>
      <pubDate>Sat, 14 Nov 2026 08:00:00 GMT</pubDate>
      <category>Pricing</category>
      <description>LLM cost per user = requests per active user per month × tokens per request × the per-1M token prices. Measure it per feature from the usage field, price tiers on the 90th-percentile user rather than the average, and add fair-use limits for the tail. Cheaper models, caching and low reasoning effort cut the number directly.</description>
    </item>
    <item>
      <title>LLM latency: TTFT vs tokens per second vs end-to-end time</title>
      <link>https://broadrouter.com/blog/llm-latency-ttft-throughput/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/llm-latency-ttft-throughput/</guid>
      <pubDate>Sat, 14 Nov 2026 08:00:00 GMT</pubDate>
      <category>Engineering</category>
      <description>Time to first token (TTFT) is the wait before the first output token. Tokens per second is how fast the answer then streams. End-to-end latency is the total: TTFT plus generation time, roughly output tokens divided by per-user speed. Prompt length and queueing drive TTFT, model size and load drive speed, and answer length drives the total. Measure all three at p50 and p95.</description>
    </item>
    <item>
      <title>LLM streaming explained: SSE chunks, deltas, usage and cancel</title>
      <link>https://broadrouter.com/blog/llm-streaming-sse-explained/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/llm-streaming-sse-explained/</guid>
      <pubDate>Sat, 14 Nov 2026 08:00:00 GMT</pubDate>
      <category>Engineering</category>
      <description>An LLM API streams a chat completion as Server-Sent Events: each data line holds a JSON chunk whose choices[0].delta carries the next piece of text. Concatenate the deltas, stop at data: [DONE], and set stream_options include_usage to get token counts in a final chunk with an empty choices list. To cancel, close the connection.</description>
    </item>
    <item>
      <title>Long context vs RAG: when 1M tokens replaces a vector DB</title>
      <link>https://broadrouter.com/blog/long-context-vs-rag/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/long-context-vs-rag/</guid>
      <pubDate>Sat, 14 Nov 2026 08:00:00 GMT</pubDate>
      <category>Engineering</category>
      <description>Long context replaces RAG when the corpus fits in the window, changes rarely and questions need many parts of it at once; caching keeps repeat queries cheap. RAG still wins when the corpus outgrows the window, updates constantly, needs per-user permissions or must answer in a second or two. Many systems retrieve with RAG and reason with long context.</description>
    </item>
    <item>
      <title>Long-running agents: timeouts, retries and checkpoints that hold</title>
      <link>https://broadrouter.com/blog/long-running-agents-timeouts-retries/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/long-running-agents-timeouts-retries/</guid>
      <pubDate>Sat, 14 Nov 2026 08:00:00 GMT</pubDate>
      <category>Agents</category>
      <description>Treat each model call as disposable and the agent's state as durable. Set a timeout per call, stream to detect stalls, retry 429, 5xx and dropped streams with bounded backoff, and give every side-effecting tool an idempotency key. Save the message list after each step so a crash resumes from the last checkpoint, not from the start.</description>
    </item>
    <item>
      <title>Multilingual speech-to-text API: 14 languages, one endpoint</title>
      <link>https://broadrouter.com/blog/multilingual-speech-to-text-api/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/multilingual-speech-to-text-api/</guid>
      <pubDate>Sat, 14 Nov 2026 08:00:00 GMT</pubDate>
      <category>Arabic AI</category>
      <description>BroadNet Transcribe is a speech-to-text model for 14 languages, including English, Spanish, French, German, Chinese, Japanese and Arabic, served on an OpenAI-compatible /v1/audio/transcriptions endpoint. It costs $0.002 per audio minute ($0.12 per hour). For mainly Arabic audio, BroadNet Transcribe Arabic is the specialised option at $0.0042 a minute.</description>
    </item>
    <item>
      <title>n8n AI Agent on a cheap LLM: setup, model choice and cost per run</title>
      <link>https://broadrouter.com/blog/n8n-ai-agent-cheap-llm/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/n8n-ai-agent-cheap-llm/</guid>
      <pubDate>Sat, 14 Nov 2026 08:00:00 GMT</pubDate>
      <category>Agents</category>
      <description>Point n8n's OpenAI credential at a low-cost OpenAI-compatible API, enter the model by ID and turn off Use Responses API. Give the AI Agent a model with tool calling, such as Qwen3.8 27B or GLM 5.3. Keep Max Iterations low, memory short and tools few, and route simple classification to a cheaper model first.</description>
    </item>
    <item>
      <title>Open-weight vs closed AI models for business: cost, control, risk</title>
      <link>https://broadrouter.com/blog/open-weight-vs-closed-models/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/open-weight-vs-closed-models/</guid>
      <pubDate>Sat, 14 Nov 2026 08:00:00 GMT</pubDate>
      <category>Comparisons</category>
      <description>Closed models still lead at the frontier, but Epoch AI estimates the best open-weight models trail them by about four months. Open-weight models win on control, privacy options, version pinning and lock-in, because many providers can serve the same weights. Closed models win when you need the very best quality on hard tasks. Many businesses use both, routed by task.</description>
    </item>
    <item>
      <title>OpenRouter vs a flat-rate plan for AI agents: which costs less?</title>
      <link>https://broadrouter.com/blog/openrouter-alternative-flat-rate/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/openrouter-alternative-flat-rate/</guid>
      <pubDate>Sat, 14 Nov 2026 08:00:00 GMT</pubDate>
      <category>Comparisons</category>
      <description>OpenRouter bills per token at provider prices, plus a fee when you buy credits. A flat-rate plan bills per session, whatever the agent reads. For an agent that runs most of the day on one model, BroadNet's unlimited DeepSeek V4.1 Flash at $60 a month per session usually costs less. For model variety, fallback or light use, OpenRouter fits better.</description>
    </item>
    <item>
      <title>Reasoning tokens: why you pay for more output than you see</title>
      <link>https://broadrouter.com/blog/reasoning-tokens-cost/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/reasoning-tokens-cost/</guid>
      <pubDate>Sat, 14 Nov 2026 08:00:00 GMT</pubDate>
      <category>Pricing</category>
      <description>A reasoning model writes hidden thinking before its answer, and those reasoning tokens are billed at the output price. A 200-token reply can carry 1,000 or more tokens of thinking. On BroadNet Inference, DeepSeek V4.1 Flash, GLM 5.3 and Qwen3.8 27B reason. Lower the cost with reasoning_effort set to low, a sensible max_tokens, and a non-reasoning model for simple steps.</description>
    </item>
    <item>
      <title>Self-hosting an LLM vs an API: the real cost and break-even</title>
      <link>https://broadrouter.com/blog/self-hosting-llm-vs-api-cost/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/self-hosting-llm-vs-api-cost/</guid>
      <pubDate>Sat, 14 Nov 2026 08:00:00 GMT</pubDate>
      <category>Comparisons</category>
      <description>A rented GPU costs the same per hour whether busy or idle, so self-hosting beats an API only at high, steady utilization. Divide the hourly price by the tokens the server really produces per hour, add engineering and on-call time, and compare with the API price per 1M tokens. For spiky or modest traffic, the API almost always wins.</description>
    </item>
    <item>
      <title>Sending sensitive data to an LLM API: redact, minimise, log less</title>
      <link>https://broadrouter.com/blog/llm-api-sensitive-data/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/llm-api-sensitive-data/</guid>
      <pubDate>Sat, 14 Nov 2026 08:00:00 GMT</pubDate>
      <category>Privacy</category>
      <description>To send sensitive data to an LLM API, first decide which fields the task needs and drop the rest. Replace names, numbers and IDs with reversible placeholders before the request, and restore them in your own code after. Use a provider with zero data retention, keep prompt content out of your own logs, and lock API keys to your servers.</description>
    </item>
    <item>
      <title>Speech-to-text API pricing compared: price per minute in 2026</title>
      <link>https://broadrouter.com/blog/speech-to-text-api-pricing-compared/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/speech-to-text-api-pricing-compared/</guid>
      <pubDate>Sat, 14 Nov 2026 08:00:00 GMT</pubDate>
      <category>Comparisons</category>
      <description>Batch speech-to-text APIs cost roughly USD 0.0007 to 0.016 per audio minute at list price in November 2026. Deepgram Nova-3 charges USD 0.0043, OpenAI's current gpt-transcribe USD 0.0045, AWS Transcribe USD 0.006 and Google V2 standard USD 0.016. BroadNet Transcribe costs $0.002 a minute in 14 languages, and BroadNet Transcribe Arabic $0.0042.</description>
    </item>
    <item>
      <title>Switching to an OpenAI-compatible API: a migration checklist</title>
      <link>https://broadrouter.com/blog/openai-compatible-api-migration-checklist/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/openai-compatible-api-migration-checklist/</guid>
      <pubDate>Sat, 14 Nov 2026 08:00:00 GMT</pubDate>
      <category>Engineering</category>
      <description>To move to an OpenAI-compatible API, change three settings (base_url, api_key and model) and then check what differs: model IDs, how reasoning tokens count toward max_tokens, how streaming reports usage, which errors to retry, rate limits and key security. For BroadNet Inference the base URL is https://inference.broadnet.ai/v1.</description>
    </item>
    <item>
      <title>Tokens per word in Arabic vs English: why Arabic costs more</title>
      <link>https://broadrouter.com/blog/tokens-per-word-arabic-english/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/tokens-per-word-arabic-english/</guid>
      <pubDate>Sat, 14 Nov 2026 08:00:00 GMT</pubDate>
      <category>Pricing</category>
      <description>There is no single figure: it depends on each model's tokenizer. In BroadRouter's measured dataset, the same Modern Standard Arabic text needs 1.31× (Qwen3.8 27B) to 1.77× (GLM 5.3) the tokens of its English translation, and 1,000 words of English prose come to about 1,290 tokens on average. Measure your own ratio from usage.prompt_tokens.</description>
    </item>
    <item>
      <title>Tool calling with open models: schemas, parallel calls, retries</title>
      <link>https://broadrouter.com/blog/tool-calling-open-models/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/tool-calling-open-models/</guid>
      <pubDate>Sat, 14 Nov 2026 08:00:00 GMT</pubDate>
      <category>Engineering</category>
      <description>Open models on OpenAI-compatible APIs use the same tools and tool_calls format as OpenAI. Keep schemas small and flat, parse arguments as JSON strings, validate them before running anything, and return errors to the model as tool results so it can retry. Cap the steps. On BroadNet, GLM 5.3 and Qwen3.8 27B list tool support.</description>
    </item>
    <item>
      <title>Use MCP servers with OpenAI-compatible models (any provider)</title>
      <link>https://broadrouter.com/blog/mcp-with-openai-compatible-models/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/mcp-with-openai-compatible-models/</guid>
      <pubDate>Sat, 14 Nov 2026 08:00:00 GMT</pubDate>
      <category>Agents</category>
      <description>MCP is model-independent. The MCP client in your app lists each server's tools, passes them to the model as ordinary function-calling tools, and runs tools/call when the model asks. Any OpenAI-compatible model with reliable tool calling can use MCP servers. Configure them in Cline, Continue, opencode or n8n, or bridge them in about 40 lines of Python.</description>
    </item>
    <item>
      <title>Use open models in coding agents: Aider, Cline, opencode, Codex</title>
      <link>https://broadrouter.com/blog/use-open-models-in-coding-agents/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/use-open-models-in-coding-agents/</guid>
      <pubDate>Sat, 14 Nov 2026 08:00:00 GMT</pubDate>
      <category>Agents</category>
      <description>Every major coding agent takes three values for a custom model: a base URL, an API key and a model ID. Aider uses OPENAI_API_BASE, Cline and Roo Code an OpenAI Compatible provider, opencode a custom provider, Continue apiBase. Codex CLI now speaks only the Responses API, so a chat-completions endpoint needs a translating proxy such as LiteLLM.</description>
    </item>
    <item>
      <title>Voice notes to text for business: WhatsApp and Telegram to tickets</title>
      <link>https://broadrouter.com/blog/voice-notes-to-text-for-business/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/voice-notes-to-text-for-business/</guid>
      <pubDate>Sat, 14 Nov 2026 08:00:00 GMT</pubDate>
      <category>Arabic AI</category>
      <description>To turn customer voice notes into tickets, download each note from the WhatsApp Cloud API or Telegram Bot API as soon as the webhook arrives, convert it to WAV, transcribe it with the right language code, then ask an LLM to extract intent, urgency and order numbers as JSON. Create the ticket with the transcript attached and reply in text.</description>
    </item>
    <item>
      <title>What "unlimited" really means in an LLM API plan</title>
      <link>https://broadrouter.com/blog/what-unlimited-llm-api-really-means/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/what-unlimited-llm-api-really-means/</guid>
      <pubDate>Sat, 14 Nov 2026 08:00:00 GMT</pubDate>
      <category>Agents</category>
      <description>"Unlimited" in an AI plan can still hide a token meter, a rolling usage window, a weekly cap, a rate limit or a concurrency limit. Ask which of these apply. BroadNet's Full Moon plan has no token meter, no 5-hour window and no weekly cap for DeepSeek V4.1 Flash; its one limit is one request at a time per session.</description>
    </item>
    <item>
      <title>What multi-agent setups really cost: per token vs flat sessions</title>
      <link>https://broadrouter.com/blog/multi-agent-cost/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/multi-agent-cost/</guid>
      <pubDate>Sat, 14 Nov 2026 08:00:00 GMT</pubDate>
      <category>Agents</category>
      <description>A multi-agent setup costs roughly the tokens of every agent added together, plus handoffs and repeated context. Anthropic measured about 15 times the tokens of a chat for its multi-agent research system, against 4 times for one agent. Per token, the bill grows with every agent; on a flat plan it grows with the agents running at once.</description>
    </item>
    <item>
      <title>When a 1M-token context window is worth it (and when not)</title>
      <link>https://broadrouter.com/blog/million-token-context-window-guide/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/million-token-context-window-guide/</guid>
      <pubDate>Sat, 14 Nov 2026 08:00:00 GMT</pubDate>
      <category>Engineering</category>
      <description>A 1M-token context window lets one request hold a whole codebase, a contract set or hours of transcripts, so the model sees everything at once instead of retrieved fragments. It pays off for cross-file reasoning and one-off analysis; for repeated lookups over a large, changing corpus, retrieval is cheaper. DeepSeek V4.1 Flash and GLM 5.3 accept 1,048,576 tokens.</description>
    </item>
    <item>
      <title>Whisper API alternative: when to switch, and how to test one</title>
      <link>https://broadrouter.com/blog/whisper-api-alternative/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/whisper-api-alternative/</guid>
      <pubDate>Sat, 14 Nov 2026 08:00:00 GMT</pubDate>
      <category>Comparisons</category>
      <description>OpenAI lists whisper-1 as deprecated, shutting down on February 26, 2027, so every Whisper API user will move to another model. Pick the replacement by word error rate on 30 to 50 of your own clips per language, then price. BroadNet Transcribe takes the same OpenAI SDK call and costs $0.002 a minute in 14 languages.</description>
    </item>
    <item>
      <title>Zero data retention LLM API: a 10-question checklist</title>
      <link>https://broadrouter.com/blog/zero-data-retention-llm-api-checklist/</link>
      <guid isPermaLink="true">https://broadrouter.com/blog/zero-data-retention-llm-api-checklist/</guid>
      <pubDate>Sat, 14 Nov 2026 08:00:00 GMT</pubDate>
      <category>Privacy</category>
      <description>A zero-data-retention LLM API processes your prompts and outputs for the request and does not store them afterwards, in logs, caches or training sets. Ask what is kept (usually content-free usage records), for how long, who runs the GPUs and how retention is checked. BroadNet Inference keeps no prompts, images, audio or answers and checks hourly that its gateway is configured to keep no content.</description>
    </item>
  </channel>
</rss>
