Mistral Large 4 vs Mistral Large 3: The Upgrade That Costs 40% More

The deprecation notice is one sentence long, and it was the most consequential thing Mistral published in October. "This model is deprecated. Use Mistral Large 4 for new integrations," reads the documentation page for Mistral Large 3, dated 31 October 2026 — two days after the weights of Mistral Large 4 landed. Large 3 is not being withdrawn tomorrow; it is being closed to new work, which in practice means every new integration, eval and benchmark run from here points at a model that costs 40% more per token and ships under a licence that no longer lets you resell it freely. That is the trade the upgrade asks you to make, and it is not the trade the benchmark coverage describes.

What the deprecation actually says

Mistral's model page lists Mistral Large 3 at 675 billion total parameters and 41 billion active, with a 256K context, released 2 December 2025 under Apache 2.0. Its replacement, Mistral Large 4, is 1.05 trillion total and 52 billion active, with a 1M context, previewed 6 October 2026, and open-weighted on 29 October. Mistral names mistral-large-4 as the successor by API id.

Two details in that pairing deserve more attention than they got. The first is that the licence changed. Apache 2.0 became the Mistral Ecosystem Open-Weight License, a modified MIT that permits commercial use but requires a separate commercial agreement for anyone offering the model as a model-as-a-service. If you fine-tune Large 3 and sell access to the result, Large 4 asks you to talk to Mistral first. The second is that Mistral's own documentation still labels Mistral Large 4 a "Public Preview" while the licence block on the same page describes it as open-weight. Both can be true — the model is published, the release is not formally closed — but a team planning a migration should not read "preview" and "open weight" as the same maturity signal.

The price went up, and the per-task cost went down

Both statements are correct, which is what makes this upgrade confusing.

• Input tokens — $0.50 per million on Mistral Large 3, $0.70 on Mistral Large 4. Up 40%.

• Output tokens — $1.50 per million on Mistral Large 3, $2.10 on Mistral Large 4. Up 40%.

• Cached input — $0.07 per million on Mistral Large 4; Mistral's Large 3 documentation page does not list a cached rate.

• Cost per completed task on the Artificial Analysis index — $2.06 on Mistral Large 4, down 63% from the October preview's $5.54.

The 63% figure is real and it describes something different from what a buyer experiences. It compares Large 4 to its own earlier preview, when the model was both more expensive per token and far more verbose. What it does not do is compare Large 4 to Large 3, because Artificial Analysis currently publishes no Intelligence Index score for Large 3 at all — the page lists the index as not applicable and the model as having no publicly benchmarked API configuration. So the widely circulated "9 to 38.4" jump across the two models is not a measurement of this upgrade. It is a score from an index build that has since been replaced.

The score question, honestly

On the current Artificial Analysis Intelligence Index v5 — rebuilt on 27 October 2026 — the only like-for-like pair available is inside Large 4's own family. The October preview scored 25, and the released checkpoint scores 26 at high reasoning and 23 at medium. That is the measured improvement from the three extra weeks of reinforcement learning Mistral describes, and it is one point.

For the generational claim, you are left with Mistral's own evaluations: 41% on Terminal-Bench 4.0 against 28% at preview, 61.7% on DeepSWE v1.1, 79% on CyberGym-E2E, 84.5% on BrowseComp. All are vendor-reported, and the Terminal-Bench number in particular measures a different thing depending on the scaffold: 24.7% when the same weights run on Artificial Analysis's neutral mini-swe-agent harness. There is no equivalent Mistral-published suite for Large 3 to put beside it. Anyone who tells you Large 4 is exactly N points better than Large 3 is quoting an index that no longer exists in that form.

What can be said without a ruler: Large 4 adds a reasoning ladder Large 3 does not have, with low, medium and high effort levels where the old model had none, and high is the default. It doubles the context to 1M and, more consequentially for long jobs, it thinks before answering. Those are categorical changes, not scale changes. The cost of a categorical change is that old prompts behave differently — sometimes better, often more expensively.

The bill on a real workload

One hundred million input tokens and forty million output tokens a month, the same volume used elsewhere on this blog because it is roughly one mid-sized support assistant.

• Mistral Large 3 — 100 × $0.50 is $50, and 40 × $1.50 is $60, for $110 a month.

• Mistral Large 4 — 100 × $0.70 is $70, and 40 × $2.10 is $84, for $154 a month.

• The difference — $44 a month, or 40%, for identical token counts.

Identical token counts is the assumption to interrogate. Mistral Large 4 is a reasoning model by default, and reasoning models emit more tokens for the same request. Artificial Analysis measures 82,000 output tokens per index task on the released checkpoint — down 39% from the preview's 135,000, which tells you how much of the preview's expense was verbosity, and also that 82,000 is still a lot. Mistral separately claims the model uses 47% fewer tokens per task than GLM-5.3, which is a comparison against a competitor, not against Large 3.

The practical consequence: if you swap the model behind an unchanged prompt, your token count will rise as well as your rate, and the bill will move by more than 40%. Turn reasoning effort down to low for classification, extraction and short-form generation, and the gap narrows. Leave it at high for everything, and you are paying a reasoning premium on work that never needed it. That single setting is worth more than any migration discount.

The licence is the part to read twice

Apache 2.0 on Mistral Large 3 meant no conditions worth a lawyer's time. The Mistral Ecosystem Open-Weight License on Large 4 is a modified MIT: use, modify, fine-tune and commercialise freely, with a separate commercial licence required for model-as-a-service. Mistral's announcement is explicit that providers must obtain one.

The carve-outs matter for intermediaries. The definition excludes embedded end-user products and the "mere relaying of requests to models hosted by others," which leaves room for a gateway passing traffic to Mistral-hosted inference. What it does not clearly cover is hosting the weights yourself and letting other people call them — the thing a platform team does when it stands up a shared inference endpoint for its own product groups. If that is your architecture, the upgrade changes your legal position as well as your bill, and the answer is a conversation with Mistral rather than a reading of the licence.

This is also where a routing layer earns its place, with a caveat. OrcaRouter passes provider list price through at 0% markup, so a vendor price change is live on our side the same day rather than at the next contract renewal — useful when a model moves from $1.36 to $0.70 in three weeks, as this one did. But neither Mistral Large 4 nor Mistral Large 3 is in our catalogue, so we cannot route this migration for you. What we can do is keep the rest of your stack on one key while you decide, with automatic failover and 200-plus models behind a single endpoint.

A migration checklist that fits on one screen

• API id — mistral-large-2512 becomes mistral-large-4 or its mistral-large-4-0 alias; the old id keeps serving existing traffic but is closed to new integrations.

• Reasoning parameter — Large 4 exposes effort at low, medium and high, with high as the default and low replacing the old "none" mode; prompts that never set it will now reason.

• Context — 256K becomes 1M, which changes chunking strategy more than it changes the model; retrieval pipelines tuned to 256K will silently start sending much larger prompts if you let them.

• Output budget — Mistral does not publish a maximum output figure for Large 4; measure it against the endpoint you call before you assume a number.

• Vision — Mistral publishes a 1.6B vision encoder for Large 4 and does not break the encoder out for Large 3; the multimodal path is native in both, but re-run image tasks rather than assuming parity.

• Licence — if you resell inference, get the commercial agreement before you ship, not after.

• Evals — re-baseline on your own tasks at high and low effort. The vendor numbers are not transferable, and neither are your old pass rates.

Who should move now, and who should wait

Move now if your work is agentic, long-context or reasoning-heavy, and if your bill is dominated by retries rather than first-pass tokens. The 40% rate increase is not the number that matters in that profile; the number that matters is how many attempts a task takes, and the reasoning ladder is the first thing Mistral has shipped that plausibly reduces it.

Wait if your workload is high-volume, short-output and already passing. A 40% rate increase for capability you are not using is a straightforward loss, and Large 3 continues to serve. The deprecation closes it to new integrations, not to existing traffic — but treat that as a runway with an end, not a permanent option, and ask Mistral for the retirement date in writing rather than inferring it from a documentation page.

And if you are choosing between the two on the strength of a benchmark chart, check which index it came from first. The most-quoted numbers for this pair were measured on a build that was replaced on 27 October, and the only score both models share today is that Large 4 has one and Large 3 does not. A migration decision made on those grounds is a decision made on a stale ruler — which is exactly the kind of thing a 40% price rise makes expensive.