Muse Wonder vs Claude Opus 5: The Cheaper Model That Finishes the Job for Less

Meta Muse Wonder and Claude Opus 5 are listed a dollar apart on the input line and five dollars apart on the output line, and that is not where they differ in any way that matters to a budget. On Artificial Analysis's own harness — the same ten evaluations run for both, at max effort, on the same hardware — Muse Wonder finishes an Intelligence Index task for about $4.52 and Claude Opus 5 for about $5.86, a gap of 23% in Muse Wonder's favour that no rate card predicts. Meta Superintelligence Labs released Muse Wonder on October 12, 2026; the same lab released Claude Opus 5 on July 24, 2026. Both carry a million-token window, both are closed, and both are priced in the band Artificial Analysis calls expensive. The reason the bill comes out different is not price at all.

Two rate cards, and the third line that decides it

Start with what each vendor publishes, because on paper the two are close enough that a reader would reasonably call it a wash.

• Input — Muse Wonder $4.00 per million tokens; Claude Opus 5 $5.00. Muse Wonder is 20% cheaper.

• Output — Muse Wonder $20.00 per million tokens; Claude Opus 5 $25.00. Again 20%.

• Cached input — Muse Wonder $0.20 per million; Claude Opus 5 $0.50. Here Muse Wonder is 60% cheaper, and this is the line that matters most in agentic use, because a long session re-sends the same system prompt and tool schema on every turn. If 70% of your input is cache hits, the effective input price is $0.54 per million on Opus 5 and $0.29 on Muse Wonder — not a 20% gap but close to half.

• Context window — a million tokens on both. Meta's checkpoint documents a max output of 943,718 tokens against a 1,048,576-token window; the Claude rate card publishes 128,000 as the maximum output for Opus 5 inside the same million-token window. If your workload emits very long single responses rather than long reasoning traces, that difference is the whole comparison.

• Reasoning control — Muse Wonder runs reasoning mandatorily, with an effort ladder of max, xhigh, high, medium, low and minimal and a default of medium; reasoning tokens bill as output. Claude Opus 5 uses the vendor's adaptive-thinking design, where there is no temperature or top_p to set and depth is dialled through reasoning effort, with the model allocating thinking itself.

• Modality — Muse Wonder advertises text, image, speech and video input with text output. Claude Opus 5 takes text, images and files, and returns text.

Read down that list and the reasonable conclusion is "Muse Wonder is slightly cheaper and a bit more multimodal." The measured bills say something much larger, and the reason is in a column rate cards do not have.

Where the 23% actually comes from

Artificial Analysis publishes the token volume each model burns getting through its index, and it is the number that turns two similar price sheets into two different invoices.

• Cost per Intelligence Index task — Muse Wonder about $4.52; Claude Opus 5 about $5.86. Both figures are the evaluator's, not either vendor's.

• Output tokens consumed on that run — Muse Wonder about 130M; Claude Opus 5 about 140M, against a board median near 85M.

• What the evaluator calls each — Muse Wonder "somewhat verbose"; Claude Opus 5 "very verbose".

• Speed on the same run — Muse Wonder was not the fast one here; Claude Opus 5 is measured at roughly 54 tokens per second by Artificial Analysis, described as notably slow against a board average near 84.

• Our own instrumentation, seven days to October 12 — Claude Opus 5 p50 time-to-first-token 6.1 seconds, 76.8 output tokens per second, 3.7% error rate across the window. That is OrcaRouter's playground measurement, taken from live traffic through our endpoint, and it is a different harness from the evaluator's, so the two speeds are not comparable to each other.

So the price-per-task gap is not being driven by a dramatic verbosity difference — 130M against 140M is 7%. It is the combination of a 20% lower output rate, a 60% lower cache rate and a smaller token appetite, stacked. Each of those is individually modest; together they are the difference between a model that costs about $4.50 a task and one that costs about $5.90.

What the independent scores say, including the comparison you should probably be making

On the same index, Muse Wonder (max) scores 52 and Claude Opus 5 (max) scores 51. One point, on a composite spanning a field whose median is 27, is not a tier. Treat these as the same capability class and make the decision on cost, latency and modality rather than on the score.

The complication is that Claude Opus 5 is not the newest model from that lab.

• Claude Opus 5.5 (max) — Intelligence Index 58, about $5.98 per index task, about 260M output tokens, $4.00 in and $20.00 out, roughly 96 tokens per second. Same list price as Muse Wonder, seven points higher on the index, and nearly double the token appetite that produces it.

• Claude Opus 5 (max) — 51, about $5.86 per task, about 140M tokens, $5.00 in and $25.00 out, roughly 54 tokens per second.

• Muse Wonder (max) — 52, about $4.52 per task, about 130M tokens, $4.00 in and $20.00 out.

That is the honest shape of this matchup. Against Opus 5, Muse Wonder is ahead on every axis at once: cheaper per token, cheaper per task, one point higher, cheaper on cache. Against Opus 5.5 it is seven points behind and slightly cheaper per task — a straight capability-for-money trade at an identical list price. The interesting finding is the quiet one: Claude Opus 5.5 costs more per task than Claude Opus 5 despite sharing a token price with Muse Wonder, because it generates 260M tokens where Opus 5 generates 140M. Nothing in the Opus rate card tells you that.

For Muse Wonder's own part, the two benchmarks worth naming are the ones Meta's own table leads with and the ones that make its advantage conditional. Meta reports Muse Wonder Full at 75.3 and Muse Wonder at 30.8 on JobBench, a spread of 44 points between two checkpoints of the same model — vendor-reported, produced in Meta's harness, and not reproduced by anyone. Whether that reflects a genuine capability split or a harness artefact is unresolved, and it is the single best reason to run your own tasks before committing.

Availability is the axis neither table covers

Claude Opus 5 has an API identifier, a rate card, a published deprecation floor and three surfaces to reach it on: the vendor's own platform, the major clouds, and any gateway that carries it. Muse Wonder has one commercial path — the Meta Model API, as muse-wonder-1.0, plus Meta's own Muse Code agent — and a second checkpoint, muse-wonder-full-1.0, that is not sold per token at all. Meta's own material describes Full as included with a Muse Code subscription rather than available as an API tier, on plans published at $50 a month for Power Usage and $200 a month for Maximum Usage.

That matters to the comparison in a way that has nothing to do with benchmarks. The higher-scoring of Meta's two checkpoints — 56 on the Artificial Analysis index, ranked fourth of 699 models — is the one you cannot call with an API key. The one you can call scores 52. Every spec sheet that treats "Muse Wonder" as a single model is choosing which of those two numbers to print without saying so.

How to settle this on your own workload

The argument above is built from somebody else's harness on somebody else's tasks, and the $4.52-versus-$5.86 gap is an average over ten evaluations you probably do not run. The number that should decide your choice is the cost per completed job on your own traffic, and the only way to get it is to run both.

That is the specific thing a single endpoint buys you. Claude Opus 5 is on OrcaRouter today at vendor list price — $5.00 in, $25.00 out, $0.50 per million cache reads, the million-token window and the 128,000-token output ceiling intact, served over both the GPT chat-completions format and the vendor's native Messages API, so an existing Claude integration moves across without a rewrite. Provider list prices pass through unchanged and we take no markup, which means a vendor price cut lands on our side the same day rather than at the next billing cycle.

Muse Wonder is not on OrcaRouter, and we will not describe it as though it were: there is no meta/muse-wonder entry in our catalogue. Reaching it means Meta's own Model API and Muse Code. What we do carry from the same lab is Muse Spark 1.2, Meta's multimodal reasoning checkpoint, at $1.25 in and $4.25 out with $0.15 cache reads across a 1,048,576-token window — not the same model, not the same capability class, but the cheap Meta route if what you actually want is the family rather than the flagship.

If you want the apples-to-apples test — same prompts, same effort settings, your own token counts — you can run the Claude side through one key today, and when Muse Wonder reaches a provider we route it will appear at Meta's price with the same pass-through rule. Until then the honest instruction is: measure Opus 5 on your tasks, compare the result against the per-task figures above, and treat Muse Wonder's 23% advantage as a hypothesis rather than a fact.

The verdict

Against Claude Opus 5, Muse Wonder wins on the numbers that are actually measured: one index point, 23% lower cost per completed task, a materially cheaper cache, a wider input surface and a larger maximum output. It is the better model per dollar for long-horizon work, and the gap is real rather than a rounding artefact — the two agree closely on score while disagreeing 20% on price and 60% on cache.

Against Claude Opus 5.5 the same model loses, and loses clearly: seven index points for a per-task cost that is only about 25% lower. If your workload is one where seven points on a composite index is worth 25% more per job, Muse Wonder is not the answer and neither is Opus 5.

What neither comparison settles is the checkpoint question. Two of Meta's three published Muse Wonder numbers describe a model you cannot buy by the token, and the subscription that does reach it costs more per month than most teams' entire API spend. Plan against the 52, not the 56.