Mistral's new flagship costs less than its own mid-tier — and the price on the model card is a sale with no published end date

Mistral released Mistral Large 4 into public preview on 6 October 2026 — a natively multimodal mixture-of-experts model the lab has nicknamed Le Chonk, with 1.05 trillion total parameters and a 1M-token context window. The parameter count is the headline. The price list is the story.

What shipped

Large 4 is available today as a preview API on Mistral Studio. The model card gives the version as mistral-large-4, a 1M context window and 1.05 trillion total parameters with 52 billion active, alongside a 1.6-billion-parameter vision encoder. The launch post puts active parameters at 49 billion; the two Mistral pages disagree by three billion and neither corrects the other.

On benchmarks Mistral leads with security rather than coding: top five globally on the Artificial Analysis Cyber Index, 82% on vulnerability reproduction, and 93% across 40 Cybench exercises. Elsewhere: 61.7% on DeepSWE 1.1, 59.9% on AutomationBench, 28.3% on Terminal-Bench 4.0, and 42% on the Dense 200 visual-grounding test — which Mistral notes edges GPT-6 Astra’s 41%.

Weights are promised “by the end of the month.” No licence is named.

The flagship undercuts the mid-tier

Here is the rate card next to the model Large 4 sits above, both taken from Mistral’s own documentation:

At its list price — ignoring the sale entirely — Large 4 is 9% cheaper than Medium 3.5 on input and 44% cheaper on output, with four times the context window. There is no axis on which Medium 3.5 is the better purchase.

That matters more than a normal launch discount because of what Medium 3.5 is marketed as. Mistral’s documentation describes it as the “frontier-class multimodal model optimized for agentic and coding use cases” — the tier Mistral points developers at. It was released on 28 April 2026 under a Modified MIT licence with a 256K window. Five months later the vendor’s own top model is cheaper than it on both ends.

Medium 3.5 has not been deprecated. It carries no retirement date, and Mistral’s lifecycle policy promises six months’ notice for generally available models, so nobody is being forced anywhere. But a mid-tier priced above the flagship is an unstable arrangement with two likely endings: Medium gets a price cut, or Medium gets a deprecation notice. Either way the planning action is the same — run your evaluations against Large 4 now, while the choice is still yours.

The number on the card is not the price

The second half of this is simpler and easier to get wrong.

Mistral’s model card shows $0.68 / $0.07 / $2.09 as the live rates, with $1.36 / $0.14 / $4.18 struck through beside them. The struck-through figures are the list price. Everything above is computed from the list, because the list is what you will be paying for all but the first days of this model’s life.

What Mistral has not published is when the sale ends. The launch post does not mention a discount at all. The model card shows both price sets with no date and no duration. Secondary coverage describes the arrangement as 50% off for the first two weeks from launch, which would land the reversion on or about 20 October 2026 — but that date appears on no Mistral page we could fetch, and we are reporting it as unofficial.

This is a recurring shape rather than a Mistral quirk. Google’s Gemini 3.8 Flash TTS shipped with a documented doubling on 1 January 2027 and ElevenLabs set a dated expiry on its own launch discount. Those are better practice than this: a published cliff can go in a spreadsheet. An undated sale cannot, so the only safe way to model Large 4 is to pretend the discount does not exist.

If you are building a cost case this week, the one-line rule: budget $1.36 and $4.18, and treat anything you save before the reversion as a rebate rather than a rate.

Mistral’s own pricing page is not a source

One practical note for anyone trying to verify the above, because we tried.

Do not budget from mistral.ai/pricing. The public pricing page lists no per-model rate for any generalist model. The only per-token figure on it is an illustrative FAQ example — “Mistral Large: $0.5/M tokens in and $1.5/M tokens out” — which corresponds to no model in the current lineup at any rate, list or sale. Our own price monitor flagged the page as changed this week, with $10 and $30 disappearing from it and nothing replacing them. The page is a marketing surface with an outdated worked example on it.

The authoritative rates live on the individual model cards under docs.mistral.ai/models, and those are what the table above uses. If you have a procurement process that cites a vendor pricing URL, cite the model card.

The cached-input line is worth a second look while you are there. Large 4 publishes a cached-input rate of $0.14 at list — about 10% of the input rate — and $0.07 on sale. Medium 3.5’s model card publishes no cached-input figure at all, and neither does Small 4’s. For an agent loop that re-reads a large resident context on every turn, the cached rate dominates the bill, so Large 4 is currently the only Mistral generalist model whose agent-loop cost can be calculated in advance. Mistral’s pricing page separately states that batch processing is half price and cached input saves up to 90%, but a percentage without a base rate is not something you can put in a spreadsheet.

Where it actually sits against the frontier

Large 4 is not a Claude or GPT replacement for hard agentic work, and Mistral’s own benchmark table says so. 28.3% on Terminal-Bench 4.0 against Claude Opus 5.5’s reported 66.4% is not a close contest, and Terminal-Bench is the benchmark that most resembles letting an agent run unattended in a repository.

What it is instead is the cheapest frontier-class European option by a wide margin, and the gap is wider than the rate card implies because the tokenisers differ. Per million tokens, Large 4 at list ($5.54 for a million in plus a million out) is 4.3 times cheaper than Opus 5.5 ($24). Normalise to characters — Anthropic’s current tokeniser runs about 2.69 characters per token against the 4.0 rule of thumb used for Mistral — and Large 4 lands near $1.39 per million characters each way against Opus 5.5’s $8.92, a 6.4× gap. The caveat is real: Mistral publishes no characters-per-token figure, so that normalisation rests on an assumption. Measure it on your own text before you commit. Our API cost calculator applies the tokeniser rather than the headline rate.

Combined with the things that already made Mistral the European pick — Regional Endpoints and a priced in-region Priority Tier, EU jurisdiction, genuinely first-class European languages — Large 4 strengthens a specific case: high-volume, well-specified generation where the data must stay in Europe and the task does not require an autonomous agent. It does not touch the agentic-coding leaderboard.

What to do

If you run Medium 3.5 in production. Start the Large 4 evaluation this week. The price argument for staying is gone and the retirement risk is now non-zero.

If you were waiting for cheap frontier-class European inference. This is it, at the list price. Do not build the business case on $0.68.

If your reason for Mistral is self-hosting. Wait. The weights are promised for late October with no licence named. Mistral Small 4 and the Ministral 3 series are what you can run today.

If you are choosing a coding agent. Nothing here changes that decision. See best AI coding tools and best AI chatbots for the current field, and the model retirement calendar and pricing tracker for what else moves this quarter.

Full background: the Mistral review.

Questions or corrections? Email Pick Right. Want the full list? See all news.