Open-weight reasoning models: what thinking costs, and how to choose one
A reasoning model writes a hidden chain of thought before its answer, and those thinking tokens are billed as output: the visible paragraph is only part of what you pay for. Many open-weight models now reason, from GPT-OSS 120B to GLM-5.2, and Evanesca runs several of them in France. Choose one by measuring how much it thinks on your own hard prompts, not by its place on a leaderboard.
Published 6 Oct 2026 · checked on 20 Oct 2026
What a reasoning model does differently
Before it answers, a reasoning model produces a stretch of thinking: it explores the problem, checks intermediate results, corrects itself, then writes the reply. In an OpenAI-compatible API those tokens count in completion_tokens, and some operators also return the thinking in a separate field next to the answer.
On Evanesca, completion tokens — reasoning included — are billed at the model's output price (API documentation). That is the whole economics of these models: they are worth their thinking on maths, planning, tricky code or multi-step analysis, and they waste it on summaries, extraction or simple questions.
The cost is in the thinking
Compare reasoning models on output tokens per useful answer, not on the price per million. An illustration, not a measurement: a 300-token question with a 400-token answer, then the same answer after 2,000 tokens of thinking, at Evanesca's prices.
| Model | Answer only | With 2,000 thinking tokens | Ratio |
|---|---|---|---|
| DeepSeek V4 Flash | €0.0007 | €0.0034 | × 4.6 |
| GLM-5.2 | €0.0046 | €0.0231 | × 5.0 |
The visible answer is the same in both columns; only the thinking changed. Two models with the same rates can therefore differ several times over in real cost if one thinks three times longer than the other for the same result — which is why the number to measure is the thinking a model spends on your problems.
Thinking budgets, and how to set them
The most common surprise is an empty answer. A reasoning model called with a small max_tokens can spend the whole budget thinking and stop before writing anything, with finish_reason set to “length”. Leave max_tokens out, or allow about 1,000 tokens or more.
Several makers expose a dial for how long their model thinks; their cards, read on 6 October 2026, describe it. OpenAI's card for GPT-OSS 120B offers low, medium and high reasoning effort (model card). DeepSeek's card for DeepSeek V4 Flash lists low, high and max, and recommends allowing up to 384K output tokens at the two higher levels (model card). Qwen's card says Qwen3.8 27B thinks by default and can be told not to, request by request (model card). Evanesca passes request parameters on to the operator; whether an operator applies a given one is its choice, so check the usage figures that come back.
Thinking also takes time: on a long prompt, a reasoning model can work for minutes before the answer is complete. Raise your HTTP client's timeout, or stream the answer.
The open-weight reasoning models on Evanesca
Evanesca's catalogue flags these models as reasoning, sorted by output price:
| Model | Maker | Parameters | Context | Tool calling | Output price | Run by |
|---|---|---|---|---|---|---|
| GPT-OSS 120B | OpenAI | 117B (5.1B active) | — | Yes | €1.008 | Scaleway, Paris (fallback: OVHcloud) |
| DeepSeek V4 Flash | DeepSeek | 284B (13B active) | 256k | Yes | €1.344 | Scaleway, Paris |
| Qwen3.6 35B | Qwen | 36B (3B active) | 256k | Yes | €2.52 | Scaleway, Paris |
| DeepSeek V4.1 Flash | DeepSeek | 763B | 1M | Yes | €3.024 | AKI.IO, Germany (fallback: Scaleway) |
| Qwen3.8 27B | Qwen | 27.8B | 256k | Yes | €5.544 | Scaleway, Paris (fallback: OVHcloud) |
| Qwen3.5 397B | Qwen | 397B (17B active) | 250k | Yes | €6.048 | Scaleway, Paris |
| GLM-5.3 | Zhipu / Z.ai | 753B | 512k | Yes | €9.24 | AKI.IO, Germany (fallback: Scaleway) |
| GLM-5.2 | Zhipu / Z.ai | 753B | 256k | Yes | €9.24 | Scaleway, Paris |
Mistral Large 4, Apertus v1.5 70B, Gemma 4 26B, Mistral Small 4, Qwen3 Coder 30B, Qwen3 235B, Mistral Small 3.2, Llama 3.3 70B and Pixtral 12B carry no reasoning flag: the catalogue marks only what has been established, so a missing flag means “not established”, never “no”.
Read the context window on the operator, not on the card
A model card states what the weights can do; the operator decides what it serves. Z.ai's card for GLM-5.2 announces a “solid 1M-token context” (model card); Scaleway serves it with 256k. Qwen's card for Qwen3.8 27B says its own hosted version has 1M tokens by default; Scaleway serves 256k. The catalogue prints the operator's figure; the cards were read on 6 October 2026.
This matters more for reasoning models, because the thinking fills the window too. The widest window in Evanesca's list is 1M today.
Measure on your own prompts
- Collect ten to twenty of your hardest real prompts — ones a plain chat model gets wrong.
- Run each candidate on them with the same settings.
- Score the answers yourself, for correctness rather than length.
- Note the completion tokens of each answer from the usage object.
- Divide the total cost by the number of correct answers, and repeat with a cheaper model.
Leaderboards are a starting point, not a verdict. They reorder often, reward familiarity with known test formats, and say nothing about how many tokens a model spends thinking — the figure that sets your bill. On per-token pricing, a focused evaluation costs cents, and Evanesca shows the exact cost under each answer.
Which one for which job
- Deep reasoning — hard maths, planning, difficult code: the largest models earn their thinking. Compare them on your own examples.
- Volume, or mixed work: a cheaper reasoning model that thinks less can be the better buy even if it is slightly weaker head to head.
- Agents: check the tool-calling column, then verify on the operator you will use (how to check tool calling).
- Work that must stay under European law: every model on Evanesca runs in Europe.
And many tasks need no reasoning model at all: send summaries and extraction to a plain chat model, and keep the thinkers for the problems that need them. Prices for every model are on the pricing page; the guide to open-weight models covers the rest of the catalogue.
Frequently asked questions
Which open-weight reasoning model is the best?
There is no single winner: it depends on how hard your problems are and how many thinking tokens a model spends on them. In Evanesca's catalogue, the reasoning models range from GPT-OSS 120B to GLM-5.2; measure two or three on your own hard prompts and compare the cost per correct answer.
Why do reasoning models cost more than chat models?
Because they write a hidden chain of thought before the answer, and those tokens are billed as output. A hard question can produce far more thinking than visible text, so the real cost depends on how efficiently the model thinks, not only on its price per million.
Why did a reasoning model return an empty answer?
Most often because max_tokens was too small: the model spent the whole budget thinking and stopped with finish_reason “length” before writing the answer. Leave max_tokens out, or allow about 1,000 tokens or more.
Can reasoning models run under EU jurisdiction?
Yes. Evanesca's catalogue includes several reasoning models, all run in Europe — most of them in France by Scaleway, a French company with no US parent, with OVHcloud as a fallback for some. The models page shows where each one runs.
Read next
- Open-weight AI models in 2026: which to use for what, and who runs themThe open-weight models of 2026 — GLM, Qwen, DeepSeek, GPT-OSS, Llama, Mistral — sorted by use, with size, licence, price and the company that runs them.
- What is open-weight AI, and what can you do with the weights?An open-weight model publishes its weights, the numbers it learned in training. What you get, what you don't, how much memory they need, and who can run them.
- What is open-source AI?Open-source AI lets anyone use, study, modify and share a model, with its weights, code and training-data information. What qualifies, and what it changes.