Meta Muse Wonder Review: $4/$20 Pricing, a $4.52 Task
Muse Wonder costs the same as Claude Opus 5.5 at $4/$20 but scores 52 to Opus's 58, and runs $4.52 per index task versus $0.72 for GPT-6.1 Sol. Muse Wonder Full has no API at all.
Meta shipped a frontier model — and the best half is not for sale
On October 12, 2026, Meta Superintelligence Labs released Muse Wonder and Muse Wonder Full. TechCrunch called it the company's first attempt at a frontier model. Where Muse Spark aimed at low latency, the Wonder pair is larger and aimed at coding and long-running agentic work.
Three facts matter before any benchmark table.
- Muse Wonder (model ID
muse-wonder-1.0) takes text, image, video and audio in, returns text, and carries a 1,048,576-token context. On the Meta Model API it costs $4 per million input tokens, $0.20 cached, and $20 per million output. - Muse Wonder Full, the stronger variant, is not served through the API at any price. It is available only to Muse Code subscribers.
- Both are closed-weight. The Llama era of "download it and run it yourself" ends here.
The model that scores highest in the launch materials is the one a developer cannot buy. That is not a pricing decision but an access decision, and it belongs at the top of the review rather than the end.
The rate card doubled — and landed exactly on Opus
- Muse Wonder: $4.00 input / $0.20 cached / $20.00 output per million tokens.
- Muse Spark 1.3 (the predecessor): $1.25 / $0.15 / $4.25.
- Claude Opus 5.5: $4.00 / $0.20 / $20.00 — identical to Wonder.
- Claude Sonnet 5.5, Gemini 4 Argon (introductory), GPT-6.1 Sol: $2.00 / $10.00.
Meta spent two years training the market to expect downward price pressure from it, because selling tokens was never its business model. Muse Wonder reverses that. When Meta has something near the top, it charges what the top charges.
That kills the "switch to save money" case. On the independent index Wonder scores 52 against Opus 5.5's 58 at the same token price. What shipped is a slightly weaker model with a cheaper invoice only because it writes less. Six index points at that altitude is roughly the distance between frontier and near-frontier.
The reason to onboard Wonder is vendor diversification, not cost — and that is a real reason. Meta launched its enterprise platform on September 28 with MongoDB's former CEO CJ Desai running it. It wants to be a third credible supplier at the top end. It just is not undercutting anyone to get there.
Independent scores, and the shape of the wins
Artificial Analysis at max effort on Intelligence Index v4.3 puts Claude Opus 5.5 at 58, Muse Wonder Full at 56 (tying Claude Sonnet 5.5), Claude Fable 5.1, GPT-6 Astra and Gemini 4 Argon at 53, and Muse Wonder level with GPT-6.1 Sol at 52.
Meta's own 14-row scorecard splits six rows to Wonder Full and six to Opus 5.5, and the split is consistent.
Wonder Full wins on endurance:
- SWE-Marathon (ultra-long-horizon engineering): 57.5 at xhigh effort against 46.3 for Opus 5.5 and 42.5 for GPT-6 Astra.
- DeepSWE 1.1: 79.0 for Full, 78.8 for plain Wonder, against Opus 5.5 at 74.2.
- Harvey LAB-AA (legal agent work): 13.1 for Full against 10.0 for plain Wonder.
Opus 5.5 wins on cold starts:
- ProgramBench (rebuild a program clean-room): 70.5 against 59.0 for Wonder Full.
- Terminal-Bench 4.0-AA: 59.6 against 57.1 for Full and 52.0 for plain Wonder.
- HealthBench Pro: 65.6 against 62.3.
Read that as a division of labour. Wonder is built to keep going — long sessions, many tool calls, work that survives a blocker. It is not built to nail a single cold start, and an eleven-point deficit on from-scratch program reconstruction is the clearest evidence in the table.
Two caveats Meta states itself: rival models were run "best-effort," and where no third-party score existed, Meta used the highest of its internal result, a self-reported figure, or a leaderboard entry. The SWE-Marathon row used Codex for GPT-6 Astra and Claude Code for Fable 5.1, so it is not a common-harness comparison. Treat the grid as a list of claims, not a controlled experiment.
Cost per task is where this gets uncomfortable
Artificial Analysis measures dollars per completed Intelligence Index task. Muse Wonder (max) lands at $4.52, against $1.60 for its predecessor — 2.8× the cost of the model it replaces. Input is 76% of that total ($3.43) because Wonder reads 7.69 million input tokens per task, two to seven times what Gemini 4 Argon, GPT-6.1 Sol or GPT-6 Astra consume on the same evaluation.
Line those up with their scores:
- GPT-6.1 Sol: $0.72 per task, index 52 — the same score as Wonder, at a sixth of the price.
- Gemini 4 Argon: $1.99, index 53 (introductory; doubles to $4/$20 on December 1).
- GPT-6 Astra: $3.26, index 53.
- Muse Wonder: $4.52, index 52.
- Claude Opus 5.5: $5.98, index 58.
Wonder is priced like a flagship and measured like the tier below it. And because Wonder Full has no published token price, its cost per task cannot be calculated at all — you cannot budget for the model the headlines are about.
Active Memory is the pitch — and the weaker half has it too
Meta rests the launch on Active Memory (carrying what matters about a user across conversations) and Active Focus (pursuing a goal through blockers and checking in before consequential actions). That is the right problem to attack; a stateless assistant that forgets last week is not an agent. But both headline benchmarks are internal to Meta and unreproduced.
- MemoryBench: Wonder 85.6, Wonder Full 85.5, Opus 5.5 81.5, GPT-6 Astra 81.4.
- FocusBench: Wonder Full 54.2, Astra 47.7, Opus 5.5 45.8, Wonder 44.9.
The MemoryBench leader is the cheaper API model, by 0.1. And on FocusBench — the feature Meta spent the most words on — the API model loses to both GPT-6 Astra and Opus 5.5. If you cannot subscribe to Muse Code, you do not get the version of Active Focus that was advertised.
The part procurement should read twice
Wonder Full is reachable only through Muse Code at $50/month (Power Usage) or $200/month (Maximum Usage). No token price, no API. It is a seat licence for a coding tool, not a component you can route production traffic to.
Then the data default. Meta's own help centre states that Muse Code subscriptions default to Contributor mode, in which the prompts, code, files and context you send — plus the responses — may be used to train future Meta models. You can switch it off in /settings, and doing so changes how much usage your plan grants, not the price. But the default is on, and most teams will not look.
One correction, because it is circulating widely: the "60 requests per minute" Contributor limit belongs to the Muse Spark Contributor tier (60 RPM / 2.1M TPM against 3,000 / 4M on standard). Meta's current documentation lists muse-wonder-1.0 at 3,000 RPM and 4,000,000 TPM per team in Standard mode, and Contributor mode is not available for Muse Wonder on the Model API at all. Cite the generation you are reading.
So the API side is clean: $4/$20, your data out of the training set, no subscription, 1M context, no long-context premium.
What to actually do
Meta is genuinely back in the frontier race. What it shipped on October 12 is a credible, persistent, expensive model on the API and a better one you cannot buy — which says as much about how Meta intends to sell as about what it built.
目次
- Meta shipped a frontier model — and the best half is not for sale
- The rate card doubled — and landed exactly on Opus
- Independent scores, and the shape of the wins
- Cost per task is where this gets uncomfortable
- Active Memory is the pitch — and the weaker half has it too
- The part procurement should read twice
- What to actually do