Meta's new Muse Wonder costs exactly what Claude Opus 5.5 costs per token, $4 in and $20 out per million, and scores 52 to Opus 5.5's 58 on the independent Artificial Analysis Intelligence Index. It earns a pilot on two kinds of work where Meta's own model card shows a lead, legal agent deliverables and very long coding projects. The stronger Muse Wonder Full is not on the API at all. It ships only inside Muse Code subscriptions, which sign up through an individual Meta Account and default to a mode where Meta may train on your prompts, code and files.
So the decision splits in two. Your platform team can test Muse Wonder through the API under terms that rule out training. Your security team should decide this week whether engineers may buy Muse Code on their own, because that is where Wonder Full lives and where the training default sits.
What Meta Released on October 12
Meta Superintelligence Labs shipped two models, and only the weaker one is available to build on. The announcement says Muse Wonder is available through the Meta Model API and Muse Code, while Muse Wonder Full is "available today for Muse Code subscribers." Meta's model documentation lists muse-wonder-1.0 as a preview model with a 1,048,576-token context window, and says plainly that Wonder Full is not on the Model API. Setting "max" reasoning effort on Muse Wonder does not select it.
Per VentureBeat's report, both models are proprietary with no downloadable weights, accept text, image, audio and video, and output text. Wonder Full comes with the Power Usage plan at $50 a month and the Maximum Usage plan at $200 a month.
The API price, from Meta's pricing page:
| Muse Wonder (API) | Claude Opus 5.5 | Claude Sonnet 5.5 | |
|---|---|---|---|
| Input, per 1M tokens | $4.00 | $4.00 | $2.00 |
| Cached input, per 1M | $0.20 | $0.20 | $0.10 |
| Output, per 1M tokens | $20.00 | $20.00 | $10.00 |
| AA Intelligence Index | 52 | 58 | 56 |
| AA cost per index task | $4.52 | $5.98 | $5.46 |
Anthropic's figures come from its pricing page; index scores and per-task costs come from Artificial Analysis's pages for Muse Wonder, Opus 5.5 and Sonnet 5.5, all checked on October 13, 2026. Meta also bills web search at $2.50 per 1,000 queries and caps a team at 3,000 requests and 4 million tokens per minute.
Is Muse Wonder Cheaper Than Opus 5.5 in Practice?
Per task, yes, by about a quarter, even though the token rates are identical. Artificial Analysis measured $4.52 per Intelligence Index task for Muse Wonder against $5.98 for Opus 5.5, because Wonder used 130 million output tokens to run the index where Opus 5.5 used 260 million. Half the output at the same output price is most of the gap.
That saving buys a lower score. Opus 5.5 sits at 58, ranked first of 216 models; Muse Wonder sits at 52, ranked ninth. Sonnet 5.5 scores 56 at half the token price, though its per-task cost of $5.46 is still higher than Wonder's. We found the same pattern in Sonnet 5.5's own cost per task: the token rate predicts very little about the bill.
Two caveats before anyone builds a cost model on this. Artificial Analysis records Muse Wonder's time to first token at 19.27 seconds against a 5.52-second median, which rules it out of anything interactive. And Artificial Analysis itself calls the model "somewhat verbose" relative to its peers, so the token savings against Opus may not hold on your prompts. Measure output tokens on your own workload before you price anything.
Where Muse Wonder Actually Leads
Meta's model card shows real leads on legal agent work and very long coding projects, and Opus 5.5 ahead on most of the rest. The scores below come from the Muse Wonder model card, all at max reasoning effort except where noted.
| Benchmark | Wonder Full | Wonder | GPT-6 Astra | Opus 5.5 |
|---|---|---|---|---|
| Harvey LAB-AA (legal agent) | 13.1 | 10.0 | 8.6 | 4.2 |
| SWE-Marathon (ultra-long coding) | 57.5 | 53.1 | 42.5 | 46.3 |
| DeepSWE 1.1 | 79.0 | 78.8 | 74.1 | 74.2 |
| Terminal-Bench 4.0-AA | 57.1 | 52.0 | 59.1 | 59.6 |
| ProgramBench | 59.0 | 50.0 | 52.5 | 70.5 |
| HealthBench Pro | 62.3 | 60.5 | 65.3 | 65.6 |
| Vals Finance Agent V2 | 59.1 | 56.6 | 53.5 | 58.6 |
Harvey LAB-AA is the strongest result, and the absolute numbers are low. The card describes it as Artificial Analysis's version of Harvey's Legal Agent Benchmark, where a judge panel checks each legal deliverable against task criteria and hallucination checks, and the metric is the share of tasks that pass every criterion. Muse Wonder passes 10% of tasks, Opus 5.5 passes 4.2%. That is more than twice the pass rate, and it still means nine in ten deliverables fail at least one check. It is a reason to put Wonder into a bake-off against Harvey or your current legal model, with every output still reviewed by a lawyer. Our earlier look at Mistral Large 4's Harvey LAB result covers how to read this benchmark.
SWE-Marathon carries Meta's own warning. The card states that Abundant AI's leaderboard ran GPT-6 Astra through Codex and Claude Fable 5.1 through Claude Code, "so this row is not a controlled common-harness comparison," and that Wonder Full was scored at xhigh effort. The lead may be partly the harness. DeepSWE, which uses Datacurve's third-party runs for Wonder and Opus 5.5, shows a smaller but cleaner gap: 78.8 against 74.2.
On from-scratch programming Opus 5.5 is well ahead. ProgramBench, which asks a model to rebuild a complete program in a clean-room environment with every model on the same mini-swe-agent harness, has Opus 5.5 at 70.5 and Muse Wonder at 50.0. Terminal-Bench, run by Artificial Analysis, also favors Opus.
Meta's methodology note matters for every row: competitor runs it generated itself "remain best-effort: prompts, tools, and runtime may not be optimally tuned for proprietary models." Treat the rows sourced to Artificial Analysis, Vals and Datacurve as the evidence and the internal ones as Meta's claim. Our guide to which vendor benchmark numbers belong in a deck applies here line by line.
Muse Code Trains on Your Code by Default
Muse Code subscriptions default to Contributor mode, and that is the governance problem in this launch. Meta's subscriptions page says subscriptions "default to Contributor mode," in which Meta can use your content to train future models in exchange for more usage. The covered content is the prompts, code, files and context you send, plus the responses you receive. Turning it off applies to future content; the page does not say what happens to anything already sent.
Meta Muse Code signs up through a Meta Account, and the subscriptions page describes no organizational accounts, team plans or admin controls. An engineer who wants to try Wonder Full this week will pay $50 a month, and by default every repository they open in it becomes potential training data. VentureBeat also reports that Meta's terms prohibit submitting confidential or sensitive information to Contributor services, so compliance rests on each individual subscriber.
The API is different. The model docs state that Muse Wonder on the Model API runs in Standard mode only, where "your content is not used to train Meta models." Standard mode alone does not mean zero retention. VentureBeat reports that ZDR is available to qualified Model API organizations on separate approval, and that it drops server-side conversation state, background requests, file uploads and built-in web search. If your legal pilot depends on uploaded files, ask Meta which of those you lose before you sign.
Defaults on personal accounts have caused trouble before: see how OpenAI's agents leaked training images from users who never opted out. Meta's Muse Glimmer already gave engineers a Meta model that runs without passing through your gateway.
The Case for Moving Faster
The strongest argument for Muse Wonder is cost on long agent runs. If its output stays near half of Opus 5.5's on your workload, a long-running coding or document agent at the same token rate costs materially less, and the model card's MemoryBench result (85.6 for Wonder against 81.5 for Opus 5.5) suggests it holds context across long sessions well. That benchmark is internal to Meta, so it needs your own test before it counts.
The arguments against are concrete. The API model is in preview. Meta's blog says the full safety report is coming "soon," and the model card says Meta does not rule out the unmitigated model reaching its high-risk threshold in the chemical and biological domain, which is why it ships with extra classifiers. Those classifiers can refuse legitimate work in pharma or chemicals, and you will not know the refusal rate until you test.
What to Do About It
This Week:
- Add Muse Code to your blocked or unapproved software list until security has read the subscription terms. Check expense reports for Meta subscription charges at $50 or $200 a month.
- Tell engineering managers in writing that Contributor mode is on by default and covers code and files, and that no company repository goes into a personal Muse Code session.
This Month:
- If you run a legal or contract-review agent, add
muse-wonder-1.0to the bake-off against your current model on 30 to 50 of your own matters, scored pass or fail on every criterion the way Harvey LAB does. Our legal AI comparison has the grounding checks to include. - Log output tokens per completed task for Wonder and your incumbent model on the same jobs. The per-task saving is the only reason to switch at an identical token price, so measure it directly.
- Ask your Meta account contact whether ZDR approval is available for your organization and which features it disables.
Before Renewal: If your Opus 5.5 commitment is up within two quarters, bring the per-task cost from your own Muse Wonder logs to the negotiation as a second quote, since the two list prices per token are identical.
The Bottom Line
The Muse Wonder you can contract for scores six points under Opus 5.5 at the same token price, leads on legal deliverables and very long coding runs, and runs cheaper per task because it writes less. The better model is reachable only through a personal coding subscription that trains on submitted code unless the user switches it off.
Pilot the API on legal work, and block personal Muse Code sign-ups until Meta offers a business plan with admin controls.
Continue Reading
- Claude Sonnet 5.5 Costs More Than Opus 5.5 at the Same Score
- Vendor AI Benchmarks: Only Numbers With Logs Belong in the Deck
- Mistral Large 4 Beats Kimi K3 on Legal Work With No Licence Yet
- Harvey Swapped Models After Agents Sank Its Margin to -50%
- OpenAI's Agents Leaked 53 Images From Users Who Never Opted Out
- Meta's Agent Model Fits on a Laptop. Nothing Logs It.