Mistral Large 4 Pricing: $1.36 In, $4.18 Out, and a Half-Price Promotion With No End Date

Mistral Large 4 has one of the clearest rate cards in the open-weight market and one of the least clear effective prices. The published lines are $1.36 per million input tokens, $0.14 per million cached input, and $4.18 per million output. Alongside them, Mistral's pricing page carries a launch promotion showing the same three lines struck through at $0.68, $0.07 and $2.09 — half price on every line — with no end date published anywhere we can find. So the price of an API call today is roughly half the list rate, and nobody outside Mistral can tell you when that stops being true.

That single ambiguity is worth more to a budget than the difference between Large 4 and any competitor, because a 50% swing on a line item is larger than most model-selection deltas. Everything below is built to let you plan for the day the promotion ends rather than being surprised by it.

The card, and where the confusing numbers come from

Read from Mistral's own pricing page on 29 October 2026:

• Input — $1.36 per million tokens.

• Cached input — $0.14 per million tokens, roughly a tenth of the input price.

• Output — $4.18 per million tokens.

• Launch promotion — $0.68 in, $0.07 cached, $2.09 out. No end date stated.

• Model — mistral-large-4, a 1.05-trillion-parameter sparse mixture-of-experts with 52 billion active parameters by Mistral's own card and 49 billion by Artificial Analysis's reading, plus a 1.6-billion-parameter vision encoder. Text and image in, text out.

• Context — 1M tokens on Mistral's card; 524,288 tokens on the configuration Artificial Analysis tests. That gap is the difference between a vendor ceiling and a served window, and it matters for pricing because a context you cannot actually fill is a context you will not be billed for.

One trap deserves its own sentence. Mistral's public pricing page still carries a generic example line reading "Mistral Large costs $0.5 /M tokens in and $1.5 /M tokens out." That is not the Large 4 row — it describes an older, smaller model — and it is roughly a third of the real input price and a quarter of the real output price. Anyone budgeting off the marketing page rather than the model's rate card is off by a factor of three.

What a task actually costs

Per-token prices are a poor planning unit, so here is the same card as work.

A document pass: one million input tokens and 100,000 output tokens. At list, $1.36 of input plus $0.418 of output — about $1.78. At the promotion, $0.68 plus $0.209 — about 89 cents. The same job, twice the price, depending on a date nobody has published.

The same job with caching: if 80% of that input is a reused prefix, the input side becomes 0.8 × $0.14 + 0.2 × $1.36 per million, which is $0.384 rather than $1.36. That is a 72% cut on the input line and the single most useful lever on this card — which is why the cached rate exists at a tenth of the standard one.

A long agent run: an agent that re-reads its own history on every turn is the workload this pricing punishes. Fifty turns at 200,000 tokens of context each is ten million input tokens if nothing is cached, $13.60 at list; with the prefix cached at $0.14 it falls to $1.40 plus the uncached growth. The difference between an agent that reuses context and one that rebuilds it is larger than the difference between Large 4 and a cheaper model.

Then there is the number that is not a price at all. Mistral reports 47% fewer tokens per task than a comparable competitor on Artificial Analysis's coding agent index. If that holds on your traffic, it changes the arithmetic more than four points on a leaderboard, because you pay for tokens rather than for rank — and it compounds with the promotion rather than replacing it.

The licence line that changes what the price means

The weights landed on 29 October: 54 safetensors shards, 1,060.6 GB in bf16, ungated, under a 2,632-byte file called the Mistral Ecosystem Open-Weight License that describes itself as a modified MIT licence. The grant is broad — use, copy, modify, distribute, sublicense, sell, further train, deploy, commercial use, no revenue threshold, no user cap.

What it reserves is one specific business model. Clause 2 requires anyone who "operates a Model as a Service business" to sign a separate commercial licence with Mistral first. Model-as-a-Service is defined as giving a third party access to the model for inference or post-training in a way that lets that third party exercise meaningful control over inputs, parameters or training data. Three carve-outs sit beside it, and they matter more than the restriction: a product whose model capability is embedded inside specific features is outside the definition; "mere relaying of requests to models hosted by others" is outside it; and internal use, or use through Mistral's own products, is exempt.

Read against a rate card, that produces three different buyers. Running the weights inside your own organisation is fine and the price is your hardware. Building a product that calls the model is fine and the price is the API line above. Standing up an API that hands customers raw access to the model is the one use that needs a commercial agreement — and no vendor's per-token price tells you what that agreement costs.

Calling it versus owning it

The self-hosting arithmetic is worth stating plainly because "open weights" invites a comparison that usually loses. The bf16 checkpoint is 1,060.6 GB of weights before you have cached a single token of KV state. The NVFP4 sibling is 610.1 GB and Mistral's own guidance puts it on four B200 or B300 GPUs; the vLLM recipe in the model card asks for tensor parallelism of 8 with an FP8 KV-cache dtype. For almost any team below hyperscale, four Blackwell GPUs plus the engineering time to serve a trillion-parameter MoE costs more per useful token than the discounted API does.

Self-hosting here buys control — data residency, licence certainty, the ability to fine-tune and ship derivatives — not savings. That is the honest reason to download 1.06 TB, and it is the test to apply before doing it: name the thing you are buying, because it will not be the per-token line.

What to plan for, and what to watch

Budget on the list rate and treat the promotion as upside, because that is the only version of this card you can defend to whoever signs off the spend. The promotion has no published end date, which means it can end without notice, and a plan built on $0.68 stops working on a day nobody announces.

Then watch four things, in order of how much they change the number: whether the half-price window closes and when; whether the 1M context is served or capped in practice, since billed input follows the served window; whether the NVFP4 conversion holds the published scores, which determines whether the cheap hosting path is real; and how clause 2 is enforced in edge cases, because "meaningful control" is doing a lot of work in that sentence and has no case law behind it.

Mistral Large 4 is not on OrcaRouter's catalogue today, so this is not a route we can offer. What we do provide is the general shape: one OpenAI-compatible endpoint over 224 models with provider list prices passed through at 0% markup, so a vendor price change — including a promotion ending — lands on the same day it is announced rather than at the next contract renewal. And when a model this new is unproven against your own traffic, automatic failover is the cheapest way to put it behind real requests without betting a production path on one endpoint.