BroadRouter

Open-weight vs closed model API cost: what the price data shows

By the BroadNet Inference team · 9 min read · Published · Updated

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.

“Open source is cheaper” is the usual shorthand. It is mostly true, but the details decide your bill. This post compares per-token prices for closed models (GPT, Claude, Gemini, Grok) and open-weight models (GLM 5.3, DeepSeek V4.1 Flash, Qwen3.8 27B, gpt-oss-120b, Gemma 4) using each provider’s own pricing page, then prices one realistic workload. Every third-party price was re-read from the provider’s page on 15 November 2026. BroadRouter is BroadNet’s website, and BroadNet sells some of these open models, so we show where other hosts charge less.

The rows come from our LLM API provider prices dataset, the raw data behind the LLM API price index. For the business side of the choice (control, lock-in, quality gap), read open-weight vs closed models for business.

How do closed and open-weight prices compare per token?

Standard, pay-as-you-go prices per 1M tokens, short prompts, checked 15 November 2026:

Model Type Seller Input Cached input Output
GPT-6.1 Sol Closed OpenAI USD 2.00 USD 0.10 USD 10.00
Claude Sonnet 5.5 Closed Anthropic USD 2.00 USD 0.10 USD 10.00
Gemini 4 Hydrogen Closed Google USD 1.50 USD 0.075 USD 7.50
Gemini 3.8 Flash Closed Google USD 0.75 USD 0.075 USD 3.75
Grok 4.3 Closed xAI USD 1.25 USD 0.20 USD 2.50
GPT-6 Luna Closed OpenAI USD 0.10 USD 0.01 USD 0.50
Claude Haiku 5.5 Closed Anthropic USD 0.10 USD 0.01 USD 0.50
GLM 5.3 Open-weight Z.ai (maker) USD 1.40 USD 0.26 USD 4.40
DeepSeek V4.1 Flash Open-weight DeepSeek (maker), peak USD 0.30 USD 0.006 USD 1.20
Qwen3.8 27B Open-weight Alibaba Cloud (maker) USD 0.50 not listed USD 3.00
gpt-oss-120b Open-weight Groq USD 0.15 USD 0.075 USD 0.60
Gemma 4 26B A4B Open-weight DeepInfra USD 0.07 not listed USD 0.34

Notes from the pages themselves:

  • Gemini 3.8 Flash is a promotional price “through December 31, 2026”. From 1 January 2027 Google lists USD 1.50 input, USD 0.15 cached and USD 7.50 output, double today’s rates.
  • Claude Haiku 5.5 applies to prompts up to 100,000 tokens. Longer prompts pay USD 0.50 input and USD 2.50 output.
  • GPT-6.1 Sol and GPT-6 Luna have higher rates above 272K input tokens. Grok 4.3 bills the whole request at a higher rate once the prompt reaches 200K tokens.
  • DeepSeek halves its prices off-peak: USD 0.15 input, USD 0.003 cached and USD 0.60 output outside 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays.

Three patterns stand out:

  1. The top of the closed range costs more. GPT-6.1 Sol and Claude Sonnet 5.5 output costs more than twice GLM 5.3’s output at its own maker.
  2. The bottom of the closed range does not. GPT-6 Luna and Claude Haiku 5.5 list a lower input price than gpt-oss-120b at Groq, and within a few cents of Gemma 4 at DeepInfra. Gemma’s output price is lower.
  3. Cached input is where vendors differ most. DeepSeek’s cached rate is 2% of its input rate; GLM 5.3’s at Z.ai is about 19%. Agents that resend the same context gain most from a low cache rate. See cached input tokens explained.

How much does the host change an open model’s price?

A closed model has one price list per vendor (plus licensed clouds). Open weights can be served by anyone, so the same model has many prices. Some examples, per 1M tokens, checked 15 November 2026:

Open model Host Input Cached input Output
GLM 5.3 Z.ai USD 1.40 USD 0.26 USD 4.40
GLM 5.3 Together AI USD 1.40 USD 0.26 USD 4.40
GLM 5.3 DeepInfra (promotion, 38% off) USD 0.5625 USD 0.125 USD 2.50
GLM 5.3 BroadNet Inference $0.78 $0.173 $3.46
DeepSeek V4.1 Flash DeepSeek, peak USD 0.30 USD 0.006 USD 1.20
DeepSeek V4.1 Flash Together AI USD 0.30 USD 0.006 USD 1.20
DeepSeek V4.1 Flash DeepInfra (promotion, 30% off) USD 0.14 USD 0.0042 USD 0.42
DeepSeek V4.1 Flash BroadNet Inference $0.175 $0.0055 $0.52
Qwen3.8 27B Alibaba Cloud USD 0.50 not listed USD 3.00
Qwen3.8 27B DeepInfra (promotion, 25% off) USD 0.15 USD 0.0375 USD 1.875
Qwen3.8 27B BroadNet Inference $0.175 $0.0435 $2.17
gpt-oss-120b Groq USD 0.15 USD 0.075 USD 0.60
gpt-oss-120b Together AI USD 0.15 not listed USD 0.60
gpt-oss-120b DeepInfra USD 0.037 not listed USD 0.17
Gemma 4 26B A4B DeepInfra USD 0.07 not listed USD 0.34
Gemma 4 26B A4B BroadNet Inference $0.10 not published $0.22

The spread is large: about 2.5 times on GLM 5.3 input and about 4 times on gpt-oss-120b input between the hosts above. DeepInfra’s GLM, DeepSeek and Qwen rows are labelled as promotional discounts, and the pages do not say when they end. BroadNet is below the maker’s list price on GLM 5.3 and Qwen3.8 27B and below DeepSeek’s peak rates, but DeepSeek’s off-peak input and cached rates are lower than BroadNet’s, and DeepInfra’s promotional prices are lower on all three models. The per-model price pages list every host we track, and how inference providers price open models explains why the numbers differ. Across all 119 models that several hosts sell, the median gap is 1.76 times: same model, different host.

Price is not the only difference between hosts. Check each one for quantisation (some serve lower-precision weights), context length, rate limits, data retention and whether tool calling is supported for that model.

What does a real workload cost on each?

Per-token tables hide the mix of input, cached input and output. Here is one workload: an in-app assistant, measured per active user per month, from our LLM cost per user example:

  • 0.45M new input tokens
  • 0.45M cached input tokens (a stable system prompt and history)
  • 0.3M output tokens, reasoning included

The formula is 0.45 × input + 0.45 × cached + 0.3 × output per 1M prices. Where a page lists no cached rate, we charge cached tokens at the full input price. Cache-write and cache-storage fees, which some vendors add, are left out.

Model and seller Cost per active user per month
GPT-6.1 Sol, OpenAI USD 3.95
Claude Sonnet 5.5, Anthropic USD 3.95
Gemini 4 Hydrogen, Google USD 2.96
GLM 5.3, Z.ai USD 2.07
Gemini 3.8 Flash, Google (promotional price) USD 1.50
GLM 5.3, BroadNet $1.47
Grok 4.3, xAI USD 1.40
Qwen3.8 27B, Alibaba Cloud (no cached rate listed) USD 1.35
Qwen3.8 27B, BroadNet $0.75
Qwen3.8 27B, DeepInfra (promotion) USD 0.65
DeepSeek V4.1 Flash, DeepSeek peak USD 0.50
gpt-oss-120b, Groq USD 0.28
DeepSeek V4.1 Flash, BroadNet $0.24
GPT-6 Luna, OpenAI USD 0.20
Claude Haiku 5.5, Anthropic USD 0.20
DeepSeek V4.1 Flash, DeepInfra (promotion) USD 0.19
gpt-oss-120b, DeepInfra (no cached rate listed) USD 0.08

Two caveats change how to read this. First, the models are not equal. A row lower in the table is only a saving if the model passes your test set. Second, Gemini 3.8 Flash’s line doubles to USD 2.99 on 1 January 2027 if Google’s listed price change goes ahead. To run your own numbers, use the LLM API cost calculator.

Why compare by task tier, not by model name?

“Open vs closed” is the wrong axis for a price decision. The useful question is: which is the least expensive model that does this task well enough? Tiers also move differently over time: in archived pricing pages, Claude Opus and the GPT main line got much cheaper while Claude Haiku and Gemini Flash got dearer (OpenAI vs Anthropic vs Google API price history).

  • Simple, high-volume work (classification, extraction, short summaries): small closed models and small or mid open models all sit near the bottom of the table. Pick on accuracy, latency and data terms.
  • Agents and coding: tool calling, long context and cached input dominate. Open models with a low cache rate, or a flat-rate plan, can cost a fraction of a frontier closed model.
  • The hardest reasoning: closed frontier models still tend to lead. Paying more per token can cost less per correct answer.

Measure cost per correct answer on 50 to 100 real cases. How to evaluate an LLM before switching gives a method.

What else changes the real cost?

  • Tokenizers. The same text becomes a different number of tokens on each model. In BroadRouter’s tokenizer dataset, Modern Standard Arabic needs 1.31× (Qwen3.8 27B) to 1.77× (GLM 5.3) the tokens of its English translation, so an Arabic workload can reorder the table. See tokens per word in Arabic vs English.
  • Reasoning tokens. Reasoning models bill thinking as output on both sides. Google’s page says Gemini output includes “thinking tokens”. Set reasoning effort to the lowest level that passes your tests.
  • Batch discounts. OpenAI and Anthropic list Batch API prices at half the standard rate. If your work can wait, that narrows the gap with open models. BroadNet has no Batch API; batch processing without a Batch API shows the workaround.
  • Long-context surcharges. Several closed models charge more above 100K, 200K or 272K tokens. Check where your prompts sit.
  • Self-hosting. Open weights can also run on your own GPUs, but that only pays at high, steady use. See self-hosting an LLM vs an API.

Where BroadNet Inference fits

BroadRouter is BroadNet’s own website. BroadNet Inference is a hosted open-weight API with one OpenAI-compatible base URL, https://inference.broadnet.ai/v1, for DeepSeek V4.1 Flash, GLM 5.3, Qwen3.8 27B and BroadNet’s build of Gemma 4, plus speech to text. It does not serve closed models and does not route requests between providers. DeepSeek V4.1 Flash is partner-hosted.

On the data above, BroadNet is priced below the makers’ list prices for GLM 5.3 and Qwen3.8 27B and below DeepSeek’s peak rates, and above DeepInfra’s current promotional prices and DeepSeek’s off-peak input and cached rates. It is not the lowest-priced host for every model. What it adds is a flat plan for DeepSeek V4.1 Flash ($60 a month per session, one request at a time, no token meter), zero retention of inference content, and documented tool calling and structured output on GLM 5.3 and Qwen3.8 27B. There is no Batch API and no SLA on self-serve accounts.

Gemma 4 VL 26B runs on the $5 starting credit; the other chat models need a top-up. Compare prices or read the API docs.

Questions

Are open-weight models cheaper than GPT through an API?

Often, but compare like with like. On 15 November 2026, GLM 5.3 at Z.ai listed USD 1.40 input and USD 4.40 output per 1M tokens, below GPT-6.1 Sol at USD 2.00 and USD 10.00. OpenAI's small GPT-6 Luna listed USD 0.10 and USD 0.50, which is in the same range as hosted open models.

Why does the same open model cost different amounts at different hosts?

Hosts run the same weights on different hardware, at different precision, utilisation and margin, and some run time-limited promotions. gpt-oss-120b input was USD 0.037 per 1M at DeepInfra and USD 0.15 at Groq and Together on 15 November 2026.

Do closed model APIs have batch discounts?

Yes, and so do some open-model hosts. OpenAI and Anthropic both list Batch API prices at half the standard rate for asynchronous jobs; Fireworks and Groq list 50% off batch, and Together does for selected models. BroadNet Inference has no Batch API. Cache-read prices also differ widely between vendors.

Is the cheaper model good enough?

Only your own test set can say. Closed frontier models still lead on the hardest tasks. Run 50 to 100 real cases through each candidate and compare cost per correct answer, not cost per token.

Does language change the comparison?

Yes. Each tokenizer splits text differently. In BroadRouter's tokenizer dataset, Modern Standard Arabic needs 1.31× (Qwen3.8 27B) to 1.77× (GLM 5.3) the tokens of the English text, so per-token prices understate the gap for Arabic workloads.

Can I use these open models on BroadNet's starting credit?

Gemma 4 VL 26B, yes. DeepSeek V4.1 Flash, GLM 5.3 and Qwen3.8 27B need a top-up. The $5 starting credit covers Gemma, both transcription models and Gemma hourly keys.

Read as MarkdownAsk ChatGPTAsk ClaudeAsk Perplexity