Best open-source reasoning models in 2026

What “reasoning” means here

A reasoning model — sometimes called a “thinking” model — produces a hidden chain of thought before it emits the visible answer. During that phase it explores, corrects itself, and plans. This applies across logical reasoning, math reasoning, and code reasoning tasks: the model works through the problem before it commits to an answer. In most OpenAI-compatible APIs this thinking is returned as reasoning tokens (or a separate reasoning field), counted inside completion_tokens. Many providers — including ours — bill those tokens at the output rate, because they are real output.

That one accounting fact changes everything about how you choose. A chat model is cheap because it is lazy: it writes an answer and hopes you like it. A reasoning model is expensive because it works the problem properly — and “properly” costs output tokens. For a hard math or planning question that is worth it. For “summarize this paragraph”, burning two thousand thinking tokens is a waste.

The cost is the chain of thought

The practical upshot is that reasoning models should be compared on output tokens per useful answer, not on the headline price per million. Two models can have identical rates and very different real cost, because one uses a third as many thinking tokens to reach the same answer.

At the prices we publish (our client price, margin included — the amount actually debited), output tokens dominate the bill on a reasoning workload. So the two numbers that matter are: the output price per million tokens, and the typical reasoning quotient — how many thinking tokens a model burns on your class of problem. Both are worth measuring with a real call.

What to measure, not what to believe

Benchmark leaderboards are a starting point, not a verdict. Three caveats that apply to every 2026 reasoning-model comparison:

• Reasoning tokens are not always visible in a quick test. A one-call smoke test with a small max_tokens can return almost nothing — the model spends its whole budget thinking. That is not a broken model; it is a reasoning model doing its job. Omit max_tokens or allow plenty of headroom (we advise at least ~1,000).
• Tool calling does not travel between providers. Whether a model can call tools is a property of the serving stack, not just the weights. We re-measure it on every provider we route through (a real get_weather-style call), because the same model that calls tools on one lane can be refused on another.
• Context and dimensions should be measured, not read off a spec sheet. When a provider does not expose a field, we simply do not publish a number for it.

Why to distrust leaderboards

Universal reasoning benchmarks are the least reliable input to a buying decision, for reasons that have nothing to do with anyone cheating. First, they are released and re-ordered constantly — a model that tops a leaderboard in May is ninth by July, and the same weights did not get worse, the leaderboard just moved. Second, reasoning benchmarks reward the ability to game a known test format more than the ability to think under ambiguity in production, which is a very different skill. Third, and most importantly for cost: a leaderboard score says nothing about the reasoning quotient — the ratio of thinking tokens to useful answer, which is what determines your real bill.

The reliable procedure is narrow and boring: take ten to twenty of your hardest real prompts — the ones that genuinely fail a plain chat model — run them on each candidate under identical settings, and inspect the answers. Score them yourself (correctness, not verbosity). Look at the logs for how many reasoning tokens each answer burned. That measurement, on your data, in your format, is worth more than a month of leaderboard watching. It is also cheap to do: on per-token pricing, a focused evaluation is a few cents against a prepaid balance.

The open-weight reasoning models of 2026

Here is the live open-weight reasoning lineup we serve, with the facts we have actually verified by real calls — not copied from a marketing page. It is a curated selection, not the whole market.

Model	Parameters	License	Context	Output €/M (client)
GLM-5.2 (Zhipu/Z.ai)	753B (MoE)	MIT	—	€7.70
DeepSeek V4 Flash 0731	158B total / 13B active	MIT	1M	€1.12
Kimi K3	2.8T	—	1M	€18.24*

* Kimi K3 is served as fast access · modal, not EU sovereign — the operator is under non-EU control, so we label it so and never call it sovereign. Its rate reflects its provider’s retail cost.

Two of the three are MIT-licensed and run on EU-sovereign infrastructure. The choice is not really “which is best in a vacuum” — it is which reasoning profile fits your cost and jurisdiction constraints.

How to pick for your workload

• For deep reasoning — advanced math, planning, tricky code: the largest thinking models earn their output burn. Compare them on your own hard examples, not on a leaderboard.
• For high-volume or mixed work: a cheaper thinking model whose reasoning quotient is lower can be the smarter buy even when it is a bit less capable head-to-head.
• For agentic / tool-calling workflows: verify tool calling on the actual serving lane before committing — it is the field most likely to differ in practice.
• For jurisdiction-sensitive work: if a non-EU court must never be able to reach your prompts, restrict yourself to the models labeled EU sovereign and skip the fast-access lane entirely.

Because we bill per token from a prepaid balance that stops at zero, you can run this comparison for a few euro cents — a genuine measurement of your own workload before you commit to any model. There is no subscription and no contract to exit; every model and every price is public, and the same two-line API call works across the whole catalog.

Running them on EU-sovereign infrastructure

Open-weight does not mean sovereign-automatic. The same GLM or DeepSeek weights can be served from California or from an EU operator. What decides exposure to non-EU law is who operates the lane, not which model it is. On our catalog, the reasoning models we route through EU sovereign providers are labeled sovereign; anything requiring a lane under non-EU control is labeled fast access, in the UI and in the API response before you call.

If you need reasoning models and you need the data to stay inside EU jurisdiction, that combination is exactly what the sovereign reasoning lineup exists for — open-weight, thinking, and incapable of being subpoenaed by a US court. The broader open-source model guide covers the full catalog, and the hosting-vs-API analysis tells you when to rent your own silicon instead.

FAQ