Mistral Medium 3: What It Was, and Planning Around Its Retirement

If Mistral Medium 3 is still in your configuration, it is time to deal with it. Mistral's documentation lists it as deprecated on 22 May 2026 with retirement scheduled for 31 August 2026, and names Mistral Medium 3.5 as the replacement. That date has passed.

This guide covers what Medium 3 was, why it was a sensible choice when it launched, and how to migrate without discovering the differences in production.

The pitch: most of the frontier at a fraction of the price

Mistral announced Medium 3 on 7 May 2025 under the headline "Medium is the new large". The argument was about price-to-performance rather than raw capability. Mistral claimed it scored at or above 90% of Claude Sonnet 3.7 across the benchmarks it reported, while costing roughly eight times less, and that it outperformed Llama 4 Maverick and Cohere Command A.

The price was $0.40 per million input tokens and $2 per million output. At the time, that sat well below the established frontier models, and it is still listed on OpenRouter at the same rate at the time of writing. Mistral highlighted coding and STEM as particular strengths, and listed multimodal understanding among its features.

Those comparisons were Mistral's own, run through its internal evaluation pipeline. They were credible enough to make the model worth testing, which is the most any vendor benchmark should do.

Proprietary, but deployable

Medium 3 was not an open-weight release. Mistral's announcement named Small, Large and Pixtral as its open models and did not put Medium 3 among them. What it offered instead was deployment flexibility for enterprise customers: hybrid, on-premises or in-VPC deployment, including self-hosted environments with four or more GPUs.

Mistral also offered continued pretraining, full fine-tuning and custom post-training through its applied AI services, and named early customers in financial services, energy and healthcare. That positioned it as a model for organisations that wanted private deployment without the licence obligations of running open weights themselves — but under a commercial agreement rather than a public licence.

That distinction matters now. If you self-host Medium 3 under such an agreement, its end of life is a contract question as well as an engineering one. If you call it through an API, it is purely an engineering one.

What retirement actually means for you

Deprecation and retirement are different events. Deprecation means the vendor has stopped recommending a model and named a successor; requests still work. Retirement means the endpoint stops serving the model. For Mistral's own API, both dates for Medium 3 are now in the past.

Third-party catalogues can lag behind a vendor's schedule. The OpenRouter catalogue still lists mistral-medium-3 with a 131,072-token context window at the time of writing. A listing is not a guarantee that the underlying provider will keep serving it, so a production system that depends on it is exposed to a failure it cannot schedule. Model deprecation and migration covers how to plan for these dates before they arrive.

Where to go from Medium 3

There are three realistic destinations, and they are not equivalent:

• Mistral Medium 3.5 — Mistral's named successor. A dense 128B open-weight model with a 256K window and a per-request reasoning switch, listed at $1.50/$7.50 per million tokens. Nearly four times the per-token price, with a modified MIT licence that bars companies above a $20 million monthly revenue threshold from using the weights without a separate agreement. See the Mistral Medium 3.5 guide.
• Mistral Medium 3.1 — the August 2025 point release of the same model. It is on the same deprecation schedule, so moving to it only buys time you no longer have. The Medium 3.1 guide covers what changed.
• Mistral Large 3 — a 675B mixture-of-experts model under Apache 2.0, listed at $0.50/$1.50. Not the named successor, but close to Medium 3's original price and with a permissive licence.

The choice depends on what you used Medium 3 for. If it was coding or agent work, Medium 3.5's reasoning mode and coding focus are the relevant upgrade. If it was summarisation, extraction or general assistant work at volume, Large 3 is likely closer to your old cost profile.

A migration checklist

1. Find every reference. Search configuration, environment variables and code for mistral-medium-2505, mistral-medium-3 and any unversioned mistral-medium alias, which may already point somewhere new.
2. Build a regression set. Pull thirty real requests from your logs with outputs you know were correct.
3. Run the candidates. Test Medium 3.5 and Large 3 against that set with your existing prompts, then again with prompts adjusted for each model's recommended sampling settings.
4. Compare cost per task. A model with a higher token price can still be cheaper if it needs fewer retries. Reasoning output on Medium 3.5 counts as output tokens.
5. Pin the result. Use a dated model id, not an alias, so the next change happens on your schedule. Pinning model versions explains why.

Medium 3's lasting lesson is about the shape of the market rather than the model. It showed that a mid-tier model at a fraction of frontier pricing could cover most enterprise workloads. Its retirement shows the other half: any hosted model you depend on has an end date, and the migration is cheaper when you plan it before the endpoint goes quiet.