Mistral's Leanstral 1.5 retires 30 September and its own docs disagree on the stage

Summary. Mistral's changelog announced Leanstral 1.5's release and its 30 September 2026 retirement in the same sentence, giving the free Lean 4 proof model a 92-day API life. Thirty-six days out, Mistral's model card still badges it Public Preview, the pricing page files it under Labs, and the models index has not moved it to Deprecated. Because the stage label decides whether training-data opt-out applies, and because no third-party inference provider serves the weights, the only migration path is self-hosting.

Mistral's changelog entry for 30 June 2026 does two things in one sentence. It announces the release of Leanstral 1.5, the company's Lean 4 formal-proof engineering model, under the API identifier labs-leanstral-1-5. And it states that the model "will be retired on September 30, 2026."

That is 92 days of API life, announced up front. As of today, 25 August 2026, 36 days remain.

The retirement is not the interesting part. What matters is that outside the changelog you will not find it. The model card carries no deprecation notice, the models index still lists Leanstral 1.5 as current, and the pricing page still sells it. The three surfaces do not even agree on which lifecycle stage the model occupies, and that stage determines whether your proof corpus can be kept out of training data.

Four Mistral pages, three different answers

Mistral publishes a model lifecycle policy with five stages: Labs, Public Preview, General Availability, Deprecated and Retired. Deprecated is defined precisely as "Retirement announced. Still accessible until the retirement date."

Leanstral 1.5's retirement has been announced. By Mistral's own definition it is Deprecated. Here is what each Mistral surface actually says instead.

That is four rows and three distinct stage claims. The Deprecated and retired table on the models index does contain a Leanstral entry, but it is the predecessor: version 26.03, API identifier labs-leanstral-2603. The model with a published death date is the one missing from the table.

The Labs documentation sharpens the contradiction rather than resolving it. It states that "Labs are identified by the Labs tag in the model card and the labs- prefix in the model name identifier." Leanstral 1.5 has the prefix. Its model card shows a Public Preview tag instead. By Mistral's own identification rule, the card fails the test the docs set for it.

Why the stage label is not a cosmetic problem

The lifecycle policy attaches different obligations to each stage, and two of them are commercially material.

Data collection. Labs models, per the policy, "don't support data collection opt-out." Public Preview carries no such restriction. Leanstral 1.5 is a code agent that the release blog demonstrates operating "in a raw filesystem", editing files, running bash commands and working through a repository across millions of tokens. If you point it at a private repository, whether that source and the proofs it generates can be excluded from training turns entirely on which stage governs. The Hugging Face card settles it in practice: to use the model at all you must "opt into 'Enable Labs models'" in the Mistral admin console, on a page under the privacy settings path. Operationally it is a Labs model. The docs model card does not mention that toggle.

Price. Labs models are "free of charge". Public Preview models are "priced at the same rate as General Availability models". The model card says Price: Free, which is the Labs behaviour, not the Public Preview behaviour. If the Public Preview badge were the accurate one, the model should be billed.

Silent updates. Both stages permit them. The weights behind the endpoint can change without notice, and because Public Preview models are not eligible for -latest or -major aliases, there is no pinned General Availability version to fall back to.

The identifier in the blog does not match the one in the docs

The release blog, dated 2 July 2026, closes its Get Started section by saying the model is available "as a free API endpoint" as leanstral-1-5. The changelog, the pricing page and the model card all use labs-leanstral-1-5. The lifecycle policy states that Labs models are always prefixed labs-.

A developer copying the identifier out of the blog gets the wrong string. The policy is explicit about what happens to a wrong or retired identifier: requests "fail with a 404 error".

The blog and the docs also disagree on the parameter count. The blog says "119B total and only 6B active parameters"; the docs model card and the Hugging Face card both say 6.5B activated per token. The blog contradicts itself on the FATE results too, calling them "%87 on FATE-H and 34% on FATE-X" and then, two paragraphs later, "solving 87 and 34 problems respectively". Percentages and problem counts are not the same claim. The blog is also dated two days after the docs: 2 July against 30 June.

What retirement actually costs you

Leanstral 1.5 is Apache 2.0 and the weights are on Hugging Face, so 30 September kills the endpoint, not the model. That sounds reassuring until you check who else is serving it.

The Hugging Face card for mistralai/Leanstral-1.5-119B-A6B reports: "This model isn't deployed by any Inference Provider." The predecessor mistralai/Leanstral-2603 reports the same. There is no drop-in hosted alternative to migrate to. Downloads last month were 231 for 1.5 and 76 for the predecessor, which is not the adoption curve that attracts a third-party host.

So the migration is not "switch models". It is "stand up your own inference". The Hugging Face card gives the shape of that job.

That is five rows, and each one is a line item in a capacity plan. Four-way tensor parallelism on a 119B mixture-of-experts model is not something a proof-engineering team stands up in the last week of September.

One thing did improve between versions. The predecessor's card required a vLLM nightly build and listed transformers and SGLang support as work in progress. The 1.5 card requires a numbered stable release. Self-hosting got more tractable, which is fortunate, because it is now the only path.

Context window: three numbers on one page

The docs model card advertises 256k context and 128k max output. The Hugging Face card advertises 256k in Key Features, then recommends "≤ 200k tokens" under Recommended Settings, then ships a vLLM command that serves 200,000, then ships a Mistral Vibe agent config with auto_compact_threshold = 168000.

This matters here more than usual. The blog's headline result is test-time scaling: PutnamBench solve counts climb from 44 problems at a 50k token budget to 244 at 200k, 493 at 1M and 587 at 4M. The published 587/672 figure is a 4M-token-per-attempt result. Self-host at the recommended 200k ceiling and you sit at the point on that curve where the model solved 244.

Who this hits, and how to tell if it is you

You are exposed if any of the following is true.

• Your code, CI or agent configuration contains labs-leanstral-1-5 or leanstral-1-5. Grep for both; the blog taught one identifier and the docs taught the other.

• You installed Leanstral through Mistral Vibe with /leanstall and run vibe --agent lean. That path defaults to the hosted endpoint.

• You have "Enable Labs models" switched on in your Mistral organisation privacy settings.

• You budgeted a formal-verification workstream around a zero-cost inference line. The blog's roughly $4 per PutnamBench problem, against an estimated $300 or more for Seed-Prover 1.5 at its high setting, was measured on an endpoint that will not exist in October.

What to do before 30 September

1. Pin the weights now. Pull mistralai/Leanstral-1.5-119B-A6B from Hugging Face while the release is current. Apache 2.0 means you may keep and serve it; nothing about the API retirement changes the licence.

1. Decide hosted or self-hosted this month, not next. With no inference provider serving the model, the realistic options are your own vLLM deployment or a different prover entirely.

1. Re-baseline your benchmarks at your real token budget. The headline numbers assume budgets most self-hosted deployments will not fund.

1. Audit the Labs toggle. If "Enable Labs models" is on for the training-data reasons the policy describes, decide whether it should stay on once Leanstral is gone.

1. Track the changelog, not the model card. In this case the changelog was the only surface that carried the date.

What is still unknown

Mistral has not published a successor, yet the deprecation policy says deprecation "is announced as soon as a replacement model is available". No replacement appears in the models index. Whether the card gains a deprecation banner before 30 September, which of the two stage labels gets corrected, and whether the free endpoint is extended are all open. Nothing in the published policy commits Mistral to more than the 1 month notice both Labs and Public Preview carry, and that notice has already been exceeded.

FAQ

References

1. Mistral AI, Changelog, 30 June 2026 entry announcing labs-leanstral-1-5 and its 30 September 2026 retirement.

1. Mistral AI, Leanstral 1.5 model card, showing the Public Preview badge, Apache 2.0 licence and Free price.

1. Mistral AI, Model lifecycle policy, defining the five stages, notice periods and 404 behaviour after retirement.

1. Mistral AI, Labs, stating the Labs tag and labs- prefix identification rule and the one month removal notice.

1. Mistral AI, Pricing, listing Leanstral 1.5 in the Labs section at Free for input, cached input and output.

1. Mistral AI, Models Overview, listing Leanstral 1.5 under Other specialist models and labs-leanstral-2603 in the deprecated and retired table.

1. Mistral AI, Leanstral 1.5: Proof Abundance for All, 2 July 2026, the release blog with benchmarks, cost comparisons and the leanstral-1-5 identifier.

1. Hugging Face, mistralai/Leanstral-1.5-119B-A6B, the weights repository with architecture details, vLLM serving command and inference provider status.

1. Hugging Face, mistralai/Leanstral-2603, the predecessor's card showing the nightly vLLM requirement and provider status.

1. vLLM, Release v0.24.0, the numbered release that the Leanstral 1.5 card sets as its serving floor.

_Last updated 25 August 2026. Retirement dates, model card badges and pricing classifications were verified against Mistral's published documentation on that date and can change without notice._