Muse Wonder vs Gemini 3.1 Pro: Twenty-Two Index Points, Two Release Trains, and a Scoreboard That Moved Under One of Them

Gemini 3.1 Pro scored 57 on Artificial Analysis' Intelligence Index when it launched in February 2026 and topped the board. Read today, on the current revision of that same index, it scores 29.7. Nothing about the model changed. Muse Wonder, released on October 12, 2026, reads 52 on the same board — and the comparison most people will make, 52 against 29.7, is a real gap that is also a misleading one, because a large part of it was manufactured by a re-scoring rather than by six months of model progress. The honest version of this matchup compares two shipping models, Gemini 3.1 Pro and Muse Wonder, on a ruler that has been replaced twice, and then asks what each is actually for.

One framing before the numbers. These are not competitors in the sense a typical versus page means. Gemini 3.1 Pro is the last live Pro-tier endpoint in that line, still served under a preview identifier seven months after announcement. Muse Wonder is the newest frontier release from Meta Superintelligence Labs, with a subscription-gated higher-scoring sibling. One is a long-context multimodal workhorse whose price is its argument; the other is an output-heavy agentic model whose price is the thing to think hardest about.

The ruler changed, the model did not

This is the part of the comparison that a spec sheet will not tell you, and it is why the headline gap is deceptive in both directions.

Artificial Analysis shipped Intelligence Index v4.3 in early September 2026 and revised it again since. That update replaced Terminal-Bench v2.1 with the much harder Terminal-Bench v4.0 and swapped τ³-Banking for AutomationBench-AA, an agentic workflow benchmark built on a private held-out task set. Private held-out weighting rose to 45%. Scores fell across the frontier as a result — models that lived in the sixties on the old scale landed in the low fifties. Gemini 3.1 Pro did not fall that far; it fell to under 30.

So the same unchanged checkpoint has four published readings in circulation — 57, 46, 30 and 29.7 — depending on which page you open and when. That is not a scandal, it is what a good index looks like when the tests underneath it get harder. It is, however, a reason to never compare two numbers that came from two different revisions, which is exactly what a 52-against-29.7 line invites you to do.

What each model actually is

• Gemini 3.1 Pro — a closed-weights, natively multimodal reasoning model announced on 19 February 2026 and still filed under "Preview" at the endpoint gemini-3.1-pro-preview. There is no 3.5, 3.6, 3.7 or 3.8 Pro in the line; the Flash ladder has moved five times while the Pro ladder has not moved at all.

• Muse Wonder — a closed-weights frontier model released on 12 October 2026, served as muse-wonder-1.0 on Meta's own API and inside Muse Code. A second checkpoint, Muse Wonder Full, scores higher on Meta's own table and is Muse Code subscription only.

• Context — Gemini 3.1 Pro 1,048,576 input tokens; Muse Wonder 1,048,576. Level, and level with most of the current frontier.

• Output ceiling — Gemini 3.1 Pro 65,536 tokens in a single response; Muse Wonder 943,718. This is the widest single gap between the two models and it is fourteen times over.

• Price — Gemini 3.1 Pro $2.00 per million input tokens and $12.00 per million output below a 200,000-token prompt, stepping to $4.00 and $18.00 above it, with cached input at $0.20 and $0.40. Muse Wonder is $4.00 and $20.00 flat, cached input $0.20. At short prompts it is half the price on input and 40% cheaper on output; past 200,000 tokens the two converge almost exactly.

• Modalities in — Gemini 3.1 Pro takes text, image, video, audio and files. Muse Wonder takes text, image, speech and video. Both write text only. On paper these are the same sentence, and in practice the tooling ecosystem has had seven months of tooling built around it.

• Reasoning — Gemini 3.1 Pro supports thinking-level control with thinking tokens billed at the output rate. Muse Wonder's reasoning is mandatory, with an effort ladder from minimal to max defaulting to medium, and no non-thinking path at all.

• Independent index — Gemini 3.1 Pro 29.7 on the current revision, with an earlier reading of 46 and a launch-era reading of 57; Muse Wonder 52 on the current revision.

The re-pricing boundary most teams cross by accident

Gemini 3.1 Pro's rate card has a step in it, and it is not an overage. Below 200,000 input tokens the model is $2.00 and $12.00; above that threshold the request reprices at $4.00 and $18.00. A 210,000-token request is not billed at the cheap rate for its first 200,000 tokens — the whole request moves to the higher tier.

That is a genuine oddity for a model whose entire sales pitch is a million-token context window: the window is real, but only the first fifth of it comes at the advertised price. At $4.00 and $18.00 above the threshold, Gemini 3.1 Pro is priced within a rounding error of Muse Wonder's $4.00 and $20.00, and the cache read is identical at $0.20 per million. So the "cheaper on input" line holds precisely up to the point where you use the capability you bought the model for.

The counterweight is on the input side, where the input-side economics are genuinely strong and independent measurement backs it up. Artificial Analysis records a cost of about $0.67 per index task for Gemini 3.1 Pro and roughly 67 million output tokens generated across the index run, against a median of about 90 million for models in its price tier. A model that answers in fewer tokens at a lower rate is cheaper in a way that no rate card shows directly — and it is the number to compare against Muse Wonder's roughly $4.52 per index task, which is a consequence of both a higher rate and roughly 130 million output tokens against an 85 million median.

What you are actually choosing between

The capability gap on today's ruler is real — 52 against 29.7 is twenty-two points and it is not noise — but it is a comparison between a five-week-old frontier release and a seven-month-old preview that the vendor never promoted past preview status. Choosing between them is less about which scores higher and more about which failure mode you can live with.

Gemini 3.1 Pro's limitations are structural and known. The output ceiling of 65,536 tokens rules it out of single-turn long-form generation, which is where Muse Wonder's 943,718-token budget is most obviously an advantage. Time to first token sits at around 35.7 seconds on the published measurement, which is slow for anything interactive. And the preview identifier is the real risk: a preview endpoint can be revised or retired under you without a version bump, which is why pinning exact model IDs rather than aliases is standard practice on that model.

Muse Wonder's limitations are equally structural and much newer. Reasoning is mandatory, so every call pays for thinking tokens at $20 per million with no cheap path — Gemini 3.1 Pro's thinking level can at least be dialled toward zero. Independent measurement is one evaluator's row on a two-day-old model. And the highest-scoring version of the thing you are comparing, Muse Wonder Full, is behind a $50-to-$200 monthly subscription rather than available at the per-token prices on the same page.

There is also a benchmark-integrity dimension worth naming, because this particular pairing is a good illustration of it. Meta's launch table reports Muse Wonder Full at 59.0 on the Artificial Analysis Index while Artificial Analysis itself publishes 56, and reports the callable Muse Wonder at 56 while the evaluator reads 52. Gemini 3.1 Pro's own history is the mirror image: a 57 that was a vendor-and-press headline on one revision and a 29.7 on another. The lesson is the same in both directions — a score is only meaningful with its ruler attached, and a vendor's version of someone else's index is still a vendor number.

Running both from one key

Gemini 3.1 Pro is on OrcaRouter as google/gemini-3.1-pro-preview, at the vendor's own list price and with the 200,000-token repricing boundary carried through unchanged — $2.00 in and $12.00 out below it, $4.00 and $18.00 above, cached input $0.20 and $0.40. Nothing is marked up on our side, and the 1,048,576-token window and 65,536-token output ceiling are passed through as published.

Muse Wonder is not on OrcaRouter. There is no meta/muse-wonder entry in our catalogue and we will not describe one as if there were; today the model runs on Meta's own API and inside Muse Code. What we do carry from the same lab is Muse Spark 1.2 and Muse Spark 1.1, so the Muse family is reachable from the same key while the Wonder checkpoint builds a track record.

The practical reason that matters here is the tier boundary. A model that reprices a whole request at 200,000 tokens is the kind of thing you want to measure rather than assume, because your prompt lengths decide whether you are paying $2 or $4 for the same call. OrcaRouter puts 200-plus models behind one OpenAI-compatible endpoint with provider list prices passed through unchanged — so a run of your real traffic against Gemini 3.1 Pro at both sides of that threshold, and against whatever else is in the catalogue, costs one key and an afternoon rather than a procurement cycle. When a provider we route starts serving Muse Wonder, the other half of that test becomes a model-name change.

The verdict

• Take Gemini 3.1 Pro if your work is long-document, mixed-media analysis where the input dominates and the answer is short — the $2.00 rate below 200K and the $0.67 cost per index task are hard to beat, and the seven months of ecosystem behind it is worth something a launch-week model cannot match.

• Take Muse Wonder if you are generating long outputs in one turn or feeding speech and video straight into the model. A 943,718-token output budget is not a specification detail when the alternative is stitching generations together.

• Watch the 200,000-token line before you quote a price for Gemini 3.1 Pro. Above it, the two models cost almost the same, which reverses the intuitive answer to this comparison for exactly the workloads that use the big context window.

• Treat the 22-point gap as a direction, not a measurement. It was produced by comparing two index revisions; on the earlier revision Gemini 3.1 Pro read 46, and the same unchanged checkpoint has been published at 57, 46, 30 and 29.7 at different times.

Two models, one served under a preview identifier since February and one two days old, priced nearly identically past the 200K boundary and separated on the current board by twenty-two points. Only one of those facts would survive a change of ruler — and it is not the twenty-two points.