To choose an LLM API provider, score each candidate on six areas: the monthly cost of your real workload, rate and concurrency limits, OpenAI compatibility, data retention and training terms, the models you need, and support and reliability. Then test the top two with a week of real traffic. Below: 20 questions, a scoring sheet and the test plan.
Once you have picked, the migration checklist covers the switch itself, and how to evaluate an LLM before switching covers quality testing in depth.
What will it really cost? (questions 1 to 4)
- What does my real workload cost per month? Price input, cached input and output separately; the free LLM API cost calculator does this per model. Agents resend the same context, so cached input often dominates. Our pricing explainer walks through each line.
- Are reasoning tokens billed as output? Reasoning models can spend more tokens thinking than answering.
- Is there a flat-rate option? For always-on agents, a flat plan can beat per-token billing. See the break-even math.
- Are prices machine-readable and dated? A JSON price file lets you track changes and build cost alerts.
Our open dataset of LLM API free tiers and rate limits compares what 14 providers publish about free access, starting rate limits, card requirements and training on free-tier data, with the source and date for each value.
Which limits will you hit first? (questions 5 to 7)
- What are the requests-per-minute and concurrency limits? Concurrency (requests in flight) is what stalls parallel agents and batch jobs. See rate limits and concurrency, and size your traffic against a limit with the rate limit calculator.
- Is there a token cap, rolling window or weekly quota? “Unlimited” plans often have one. Read what unlimited really means.
- What is the context window, and is it trimmed? Check the real maximum per request, and whether the provider silently cuts long prompts.
How compatible is it with your code? (questions 8 to 10)
- Is it OpenAI-compatible, and how far? A base URL change should work. Then test streaming with usage, tool calls, structured output and error codes. The migration checklist lists what breaks.
- Does it work with my tools? SDKs, agent frameworks, IDE assistants. Check the provider’s integration guides.
- Are model IDs stable? You want versioned IDs and notice before a model is retired.
What happens to your data? (questions 11 to 13)
- Are prompts and outputs stored, and for how long? Many providers keep API content for up to 30 days by default. Compare official wording in LLM API data retention compared.
- Is my data used for training? Check the default for your tier, in writing.
- Can I sign a DPA, see subprocessors and restrict keys? Use the GDPR checklist if you process EU personal data, and ask for IP allow-lists.
Does it have the models you need? (questions 14 to 16)
- Does the provider have the model sizes and capabilities I need? Vision, tool calling, reasoning, long context, speech.
- How does each model do on my tasks? Build 30 to 50 real test cases and score them. Benchmarks are a first filter; where to compare LLMs lists the leaderboards and what each measures. The evaluation guide shows how to build the set and avoid judge bias.
- Open-weight or closed? Open-weight models can be hosted by several providers, which lowers lock-in. See open-weight vs closed models.
Will it stay up, and who answers? (questions 17 to 20)
- Is there a status page and incident history?
- Is there a written SLA for production?
- Who runs the GPUs? Own hardware, a cloud, or resold capacity from another API. Each adds a party to the data and reliability path.
- How do I reach a human? Response times, channels, and whether support can see your request content.
How do you score providers against each other?
Score each question 0 (no or unknown), 1 (partly) or 2 (yes, in writing). Weight the groups for your use case:
| Group | Questions | Weight for an agent | Weight for a regulated app |
|---|---|---|---|
| Pricing | 1 to 4 | High | Medium |
| Limits | 5 to 7 | High | Medium |
| Compatibility | 8 to 10 | Medium | Medium |
| Privacy | 11 to 13 | Medium | High |
| Models | 14 to 16 | High | High |
| Support and reliability | 17 to 20 | Medium | High |
An “unknown” scores 0 for a reason: if a provider cannot answer, you cannot rely on it.
How do you test a provider in one week?
- Pick your top two providers from the scoring sheet.
- Route a share of real traffic to each through the same OpenAI SDK, with model IDs and base URLs in config.
- Log per request: model, prompt tokens, cached tokens, completion tokens, latency, status code. Never log content.
- Trigger your peak: run your batch job or parallel agents, and watch for 429s. The errors and retries guide shows correct back-off.
- Multiply the week’s tokens by about 4.3 and price the month for each provider.
- Score the outputs on your test set.
How BroadNet Inference answers the checklist
| # | BroadNet Inference |
|---|---|
| 1, 4 | Per-token prices published in pricing.json, synced 2026-11-14 |
| 3 | Flat plans for DeepSeek V4.1 Flash: $60 a month, $15 a week or $1 an hour per session |
| 5 | Pay-as-you-go: 30 requests a minute, 4 in flight; plan sessions run one request at a time |
| 6 | Plans: no token meter, no rolling window, no weekly cap |
| 7 | Up to 1,048,576 tokens on DeepSeek V4.1 Flash and GLM 5.3; the plans page says context is never trimmed |
| 8 | OpenAI-compatible at https://inference.broadnet.ai/v1 |
| 11, 12 | No prompts or outputs retained; not used to train, fine-tune or evaluate models |
| 13 | IP allow-lists up to 32 entries; subprocessor list on request; ask for a DPA |
| 14 | Four chat models plus two speech-to-text models |
| 19 | BroadNet’s privacy statement says its own and hosted open-weight models run on BroadNet-owned or leased US hardware, with partner capacity as backup; partner-served models such as DeepSeek V4.1 Flash run with US inference partners under ZDR terms |
What we did not find in BroadNet’s docs or plans pages (checked 14 November 2026): an SLA, a status page or a model deprecation policy. Ask for them if they matter to you, and keep a second provider for critical paths. The catalogue is also smaller than an aggregator’s: if you need GPT, Claude or Gemini models, BroadNet is not the provider for that traffic. The $5 starting credit is enough for step 3 of the test plan on a light workload, but of the chat models it covers only Gemma 4 VL 26B (it also covers the two Transcribe models); the other chat models need a top-up.

