Muse Wonder vs Mistral Large 4.0: Comparing a Shipped Model Against One Still in Preview

The first honest thing to say about Muse Wonder against Mistral Large 4.0 is that only one of them is finished. Meta Superintelligence Labs shipped Muse Wonder to a commercial API on October 12, 2026: $4.00 per million input tokens, $20.00 per million output, a 1,048,576-token window, and an Artificial Analysis Intelligence Index of 52. Mistral AI put Mistral Large 4.0 into public preview on October 6, 2026 — six days earlier — at $1.36 input and $4.18 output per million, with an independent index of 38.4, no released weights, no named licence, and an output ceiling Mistral has not documented. Everything below is true this week and some of it is scheduled to stop being true before November, because Mistral has said the weights arrive by the end of the month.

Fourteen points, and what they cost

The composite gap is the widest in this series, and it is the least interesting number on the page. Both figures come from Artificial Analysis on the same revision of its index, so they compare directly:

• Intelligence Index — Muse Wonder (max) 52, ninth of 219 models; Mistral Large 4 Preview 38.4, sixty-fourth of 225

• List rate — Muse Wonder $4.00 in / $0.20 cached / $20.00 out per million; Mistral Large 4.0 $1.36 in / $0.14 cached / $4.18 out

• Cost per index task — Muse Wonder about $4.52; Mistral Large 4.0 about $1.13, a quarter of it

• Output tokens on the index run — Muse Wonder about 130 million; Mistral Large 4.0 about 200 million

• Context window — Muse Wonder 1,048,576 tokens stated and served; Mistral Large 4.0 1,048,576 in Mistral's docs, 524,288 as the independent evaluator reads it

• Parameters — Muse Wonder undisclosed; Mistral Large 4.0 1.05 trillion total with 49 billion active per token plus a 1.6-billion-parameter vision encoder

• Weights — Muse Wonder proprietary, none published; Mistral Large 4.0 proprietary today, promised open by the end of October 2026

The 4× cost ratio is the real decision, and it is not softened by the index gap the way some pairings in this series are. Muse Wonder spends fewer tokens and still lands 14 points higher, so the price difference is genuinely buying capability rather than buying verbosity. But a 38.4 index score is not a bad model — it is roughly where a competent mid-tier reasoning model sits — and the question a buyer actually faces is whether fourteen points of capability is worth quadrupling the run rate on a pipeline that processes a lot of tokens.

For most high-volume work, it is not. For agentic coding and long-horizon professional tasks, Mistral Large 4.0's preview scores are weak enough that the cheaper model is not a real substitute; the vendor's own numbers put it at 28.3 on Terminal-Bench 4.0 and 49.8 on its composite Coding Agent Index, which is a different class of result from the terminal numbers Meta publishes for Muse Wonder.

The preview clause

Mistral's own documentation cannot currently agree with itself about this model, and a comparison written today inherits that.

• The context window is stated as 1 million tokens in Mistral's docs and read as 524,288 by Artificial Analysis — a factor of two, unresolved.

• The active-parameter count appears as 49 billion on some Mistral surfaces and 52 billion on others.

• No licence is named. Mistral's own material tags the release "Open" without specifying, and the evaluator flags it proprietary with no licence field populated. Mistral Large 3 carried Apache 2.0; this one does not yet carry anything.

• The maximum output length is undocumented.

• The launch promo price — $0.68 input and $2.09 output, half the standard rate — is struck through on Mistral's pricing, which means it ends.

None of those is a scandal. They are what a preview looks like from the outside: a model whose serving configuration is still moving, sold with less notice, subject to being re-priced or retired on a shorter clock than a general-availability release. Muse Wonder has its own unresolved questions — the subscription-only Full checkpoint, the Active Memory feature nobody has tested in production — but its rate card, context window and endpoints are settled facts rather than a current best reading.

The practical consequence for anyone benchmarking these two against each other is that the Mistral column will move. A preview model's scores change between revisions of the same harness without the vendor announcing anything, and a workload tuned to a preview number should be treated as tuned to a moving target.

Where Mistral Large 4.0 has something Muse Wonder does not

Two things, and both are specific rather than incremental.

The first is sovereign deployment, and it is a procurement answer rather than a benchmark one. Mistral trains and serves this model in its own European datacentres on European hardware — the launch material cites 3,800 NVIDIA Grace Blackwell GPUs and a €3 billion Series D — and it is natively multilingual across roughly 160 languages including every official EU language. For a European public-sector or regulated buyer, that is frequently the whole decision, and Muse Wonder does not participate in it at any price. Meta's model is served by Meta, on Meta's terms.

The second is security work, where Mistral's claims are unusually concrete. The vendor reports 82% on reproducing and patching a real open-source vulnerability, which it describes as the highest of any model tested, and 93% across a 40-challenge cybersecurity suite. It further claims that Claude Opus 5.5 and GPT-6 Astra score near zero on the same test because they refuse the task. Those are vendor-reported figures about Mistral's own model and the refusal claim is a vendor's characterisation of two competitors' behaviour — treat both as claims. If the numbers hold even approximately, though, there is a class of defensive security and patch-generation work where a 38.4 index model is more useful than a 52 one, which is the kind of inversion a composite score cannot show.

On coding quality as judged by humans rather than a harness, a blind evaluation by Surge AI with labels hidden and professional annotators scored Mistral Large 4.0 at 3.74 on a five-point scale, second of five models tested — ahead of Kimi K3 at 3.59 and behind Claude Opus 5 at 4.22. That is a small, real signal that the composite understates the model on code, and it is the sort of number vendor tables never contain.

The European question, and the routing question

Mistral Large 4.0 is not in the OrcaRouter catalogue. There is no mistral route for this model on our side and this page will not imply one — Mistral's own platform, its Studio API and several third-party platforms are where the preview runs today. Meta's Muse Wonder is likewise not ours: the Meta Model API and Muse Code are where it runs, and there is no meta/muse-wonder entry in our catalogue.

What is routable here is the layer underneath both decisions. Muse Spark 1.2 — the larger Meta checkpoint this family descends from — sits in the catalogue at $1.25 input and $4.25 output per million tokens across a 1,048,576-token window with $0.15 cache reads. And the reason to hold both of these comparisons on one endpoint is that neither is a settled purchase: Mistral's preview will move when the weights land and the licence is named, and Meta's Full checkpoint may or may not get an API path. Provider list prices pass through at 0% markup, so when a vendor repriced a preview — as Mistral's struck-through launch rate will eventually do — the number on our side moves the same day rather than at the next billing cycle. The routing DSL and automatic failover exist for exactly the shape this comparison has: a workload you want to try on an unsettled model without betting a production path on it.

Choosing between a finished model and an unfinished one

• Take Mistral Large 4.0 when the deployment has to be European, when multilingual coverage in EU languages matters, or when the work is defensive security and patch generation. In those three cases the index gap is the wrong instrument.

• Take Mistral Large 4.0 when cost per task is the constraint. At $1.13 per index task against $4.52 it is the cheapest quarter in this series, and for high-volume classification, extraction and summarisation it is not close.

• Do not build on Mistral Large 4.0's preview numbers as though they were final. A model whose context window and parameter count are still disputed across the vendor's own pages is a model whose benchmark column will change.

• Take Muse Wonder when the work is agentic and long-horizon, when speech or video is the input, or when the output needs to run long — Meta's 943,718-token single-response ceiling has no counterpart here, since Mistral has not documented one.

• Take Muse Wonder when you need the capability now and can pay quadruple for it. A fourteen-point index gap on a GA endpoint with a settled rate card is a premium for certainty as much as for score.

The reason this pairing resists a clean verdict is that the two models are not competing for the same purchase. One is a frontier model with a published rate card and an unfinished feature set; the other is a cheaper, sovereign, still-forming model with a licence that does not exist yet. In three weeks Mistral Large 4.0 may be the more interesting of the two — open weights under a named licence would change every paragraph above. Right now, only one of them is something you can commit to.