Mistral Large 4 vs Claude Fable 5.1: Seventeen Index Points and a Twelve-Fold Output Price

The question in this matchup changed on 29 October 2026, and not because Mistral Large 4 got smarter. On that afternoon the model stopped being a rented preview and became a 1.06 TB download, which means the comparison with Claude Fable 5.1 is no longer only about which model answers better — it is also about which constraint you can live with. Claude Fable 5.1 is a $10-per-million-input model you can call over an API and can never own, fine-tune, or run on your own metal. Mistral Large 4 is a trillion-parameter mixture of experts you can now own, adapt and deploy inside your own network, under a licence that explicitly forbids you from reselling it as a service. One is a contract. The other is a download with a clause attached.

On capability the answer is not close, and it is worth saying that before the price discussion starts. Artificial Analysis's v5 Intelligence Index puts Claude Fable 5.1 at 43 and Mistral Large 4 at 26. That is a seventeen-point gap on the same instrument, and it survives every reasonable caveat about index versions, because both figures are v5 readings — Fable 5.1's from the live leaderboard, Large 4's from the figure Mistral quotes on its own model card. If the index is re-run, both numbers move; the ordering does not.

Where the gap actually shows up

The public benchmarks agree with the index, which is unusual. Both sets of figures below are vendor-run and neither has been independently reproduced, so treat them as claims of equal status rather than as evidence of different quality.

• Terminal-Bench 4.0 — Claude Fable 5.1 at 55.8% per Anthropic; Mistral Large 4 at 41% per Mistral. Both vendors ran their own harness, and Anthropic's own reported figure has appeared as 58.1% in some write-ups, so the exact decimal is soft in both columns.

• Long-horizon science — Claude Fable 5.1 at 52.6% on Terminal-Bench-Science 0.1, Anthropic's standout claim and one that a second lab's table has since put much lower for the same model. Mistral Large 4 has no comparable published row, which is itself informative: the model is positioned for coding and security, not for research-grade agentic science.

• Humanity's Last Exam — 60.9% without tools and 65.0% with them, for Claude Fable 5.1. Mistral published 35.0% on the same benchmark for the October preview. That preview number predates three weeks of RL and the final checkpoint, so it understates the released model — but it is the only figure either side has put on HLE.

• Security work — here Mistral Large 4 is the stronger claim, at 79% on CyberGym-E2E and 97.1% on its B3 agent-security suite after RL. Anthropic's models refuse more of this work rather than fail it, which is a policy difference as much as a capability one and should not be read as a capability gap.

• Token efficiency — Mistral's own framing is that the released model needs 47% fewer tokens per task than GLM-5.3 at a tied 48% on Artificial Analysis's Coding Agent Index. That is a rare row where a vendor publishes the cost side of a benchmark alongside the score, and it is the most useful number in either column.

A number that looks like a regression and is not

The preview version of Mistral Large 4 scored 38.4 on Artificial Analysis's index. The released version scores 26. A casual reader comparing those two figures will conclude that three weeks of reinforcement learning made the model worse by twelve points. It did not. The 38.4 was recorded on the v4.3 index; Artificial Analysis rebuilt its index in late October, and the v5 rebuild is a harsher instrument across the board — models that were in the fifties on v4.3 sit in the forties on v5. The two numbers are not measurements of the same thing, and subtracting them is the single most common error in coverage of this release.

The same warning applies to the 41% Terminal-Bench figure. Mistral reports 28.3% before the extra RL and 41% after, which is a genuine like-for-like gain on one harness. But 41% is still a low absolute score next to Claude Fable 5.1's 55.8%, and no amount of improvement framing changes that ordering either.

The licence clause that decides more than the gap

Mistral Large 4 ships under the Mistral Ecosystem Open-Weight License, described by the company as a modified MIT text. In practice it grants four things and withholds one:

• Commercial use, without a revenue threshold — there is no "you must be under $X million to use this" line.

• Fine-tuning and derivative models, which is the clause that makes the day-zero RL work possible at all.

• Internal deployment and on-premises hosting, which is the sovereign-AI pitch Mistral has built the entire release around.

• Embedding inside a product, with carve-outs covering embedded use and mere relaying of outputs.

• Not granted: operating a model-as-a-service business on it. If you want to stand up a per-token API endpoint and sell access, you need a separate commercial agreement with Mistral.

Claude Fable 5.1 has no equivalent question because it has no equivalent grant. You get an API, terms of service and a rate limit. You do not get weights, you cannot fine-tune, and you cannot put the model inside an air-gapped network at any price. Anthropic's terms add restrictions of their own that cut the same way: on the vendor's own API, zero data retention has to be specifically authorised, forced tool use is unsupported, and thinking cannot be switched off. Routing the same model through a provider with a zero-retention agreement is how teams work around the first of those. Those restrictions matter to regulated buyers in a way that a seventeen-point index gap does not.

The cost, with the verbosity caveat

List prices first, because promotional prices expire and list prices are the ones in the contract.

• Claude Fable 5.1 — $10.00 per million input tokens, $50.00 per million output, with cache reads cut to $0.25 per million. Batch pricing halves both to $5 and $25.

• Mistral Large 4 — $1.36 per million input, $4.18 per million output at list, currently discounted by roughly half to about $0.68 and $2.09 with no published end date.

At list, that is 7.4× on input and 12× on output. On the current promotion it is closer to 15× and 24×. Both are large enough that the price question outranks the benchmark question for anyone with real volume, which is the honest reason this matchup gets written at all.

The caveat is verbosity, and it is a real one. On the preview's Artificial Analysis index run, Mistral Large 4 generated output in the region of 200 million tokens against a board median near 85 million. Per-token pricing multiplies by that factor. Push the same 200 million output tokens through both models and the arithmetic is $418 for Mistral Large 4 at the promotional rate against $10,000 for Claude Fable 5.1 at list. The multiple survives the verbosity — it shrinks, it does not invert — but anyone quoting a per-token ratio without checking how many tokens the model actually emits is quoting half a number. Whether the released checkpoint is as verbose as the preview is unmeasured; the 30% output-token reduction Mistral claims from its extra RL is the company's own answer to that question.

Where each one can actually be called

This is the part of the comparison where the two models stop being symmetric. Claude Fable 5.1 is in the OrcaRouter catalogue as anthropic/claude-fable-5.1, with its 1M-token context, $10/$50 pricing and the reduced cache-read rate carried through at provider list price — no markup on top, so a vendor price change lands the same day. Mistral Large 4 is not in that catalogue. We do not serve it, and this piece is not a route to it: the released model lives on Hugging Face and behind the vendor's own API. Saying so plainly is more useful than implying otherwise.

What that means in practice depends on which side of the licence you sit. If you want a model you can run on your own hardware, audit end to end, and adapt with a LoRA adapter, only one of these two is available to you, and the seventeen-point gap is the price of that control. If you want the strongest answer per call and never intend to own anything, only one of these two is a serious option, and the twelve-fold output price is the cost of that capability. Putting both behind a single key with failover configured per route is the operational version of the same decision: it lets you default to the expensive model and fall back to a cheaper one without a second contract.

Which to pick

Take Claude Fable 5.1 if your work is long-horizon agentic execution, terminal-heavy coding, or research-grade reasoning where a wrong answer costs more than a token does. Take the 55.8% Terminal-Bench claim with the caveat that it is vendor-run, but take the ordering seriously: nothing Mistral has published puts Mistral Large 4 in the same band.

Take Mistral Large 4 if the constraint is legal or geographical rather than intellectual. A European enterprise that cannot send prompts across a border, a security team that needs the refusal policy out of the loop, a product that has to ship a model inside its own binary — none of those are solved by a better index score, and all of them are solved by 610 GB of ungated NVFP4 weights and a licence that permits internal fine-tuning.

Take neither, and route between them, if the honest answer is that you have two workloads with different constraints. That is the case the price gap is designed to create, and it is the one most teams will land in.

The open question

The number that would change this comparison is not on either vendor's slide. It is an independent re-run of the released checkpoint on Artificial Analysis's v5 index, because the 26 currently comes from Mistral quoting a third party rather than from a fresh measurement of the final weights. If three weeks of RL and 30% fewer output tokens move that figure materially, the gap narrows and the price argument gets stronger. If it does not, the released model is the preview with better manners, and the seventeen points are structural.

Until that run exists, the defensible position is the boring one: Claude Fable 5.1 is the better model, Mistral Large 4 is the better licence, and the only people for whom this is a close call are the ones who need both.