Two models, one of them rationed
Muse Wonder is the version Meta will sell you. It takes text, image, audio and video input, holds about a million tokens of context, and returns text only. It is pitched at long agent work rather than quick chat: retaining useful context across conversations, which Meta calls Active Memory, and persisting through failed attempts on a hard task, which it calls Active Focus. A million-token window is a capacity figure, not proof of reliable recall across all of it — Active Memory and Active Focus remain Meta's claims until someone tests them independently.
Muse Wonder Full is the version in the headline charts. It is limited to Muse Code subscribers, with no API price. That split matters for how you read every number that follows. The model an enterprise can actually embed in a product is standard Wonder. The model that ties or beats rivals in several of Meta's chosen tests is Full.
Read the benchmark table in two columns
On Artificial Analysis's independently run Intelligence Index v4.3, Wonder Full scores 56 and standard Wonder 52, with Opus 5.5 ahead at 58 and GPT-6 Astra at 53. In plain terms: Full is close to the front, not in front of it, and the API model sits a point behind Astra. The gains are uneven: Wonder Full trails Opus 5.5 and GPT-6 Astra on Terminal-Bench 4.0, while Artificial Analysis identifies knowledge performance as a relative weakness for both Wonder models.
Meta's own table looks stronger, especially on long-running coding and some professional agent tasks. It should be read with the caveat printed in its own methodology: the comparison mixes evaluations Meta ran itself, scores supplied by other vendors, and third-party results, and some coding comparisons used different tools for different models. That does not make the results false. It makes them a vendor table, not a controlled trial. Teams considering Wonder should benchmark the API model on their own workloads before migrating.
Cost per task cuts the other way against a simple price comparison. The token rate matches Opus 5.5 exactly, but Artificial Analysis estimates an average Intelligence Index task at $4.52 on Wonder, against $5.98 for Opus 5.5 and $3.26 for GPT-6 Astra, because models consume different amounts of tokens to finish the same work. A matching list price is not a matching bill. These are benchmark-run estimates, not a prediction of what an individual production workload will cost.
The closed turn, and the data default
Meta built much of its recent AI strategy around downloadable Llama models that developers could run on their own infrastructure. Wonder reverses that at the top end. No weights, no self-hosting, no private deployment of the frontier tier. A smaller Muse model was released openly under Apache 2.0; the frontier one was not. If you built a roadmap assuming Meta's best model would eventually be downloadable, that assumption needs revising today, not at the next Llama launch.
The second item to check before any trial is which service mode you are in. Muse Wonder's Model API is available in Standard mode only, where Meta says submitted content is not used to train its models; Contributor mode is not offered for Wonder through the API. Muse Code subscriptions, the only door to Wonder Full, default to Contributor mode, where submitted code and related material may be used to improve future models; users can turn it off in settings. Turning it off changes the treatment of subsequent submissions; developers should not assume it withdraws permission already granted for earlier material. A developer who pastes proprietary code into Muse Code on day one, before touching that setting, has already made the decision they thought they were still considering. And "not used for training" is not the same as "never retained": zero-data-retention is a separate arrangement, available to qualified organisations with feature trade-offs.
What to do with this
Treat Wonder as a serious new candidate for agent and coding workloads, priced at the frontier, and test the API model — the one you can deploy — rather than anchoring on Full's scores. Keep confidential code out of Muse Code until Contributor mode is off. And do not plan around an open-weight Wonder until Meta says there is one.
Sources
- [1] TechCrunch, October 12, 2026 — launch coverageRead source
- [2] VentureBeat, October 12, 2026 — benchmark table, API pricing, Muse Code tiers, data modesRead source
- [3] Artificial Analysis Intelligence Index v4.3 scores and task-cost estimates, as reported in the coverage above; pricing and data-mode details per Meta's model card and pricing documentation