15 OCT 2026 — Meta released Muse Wonder on 12 October, its first model built to compete at the frontier with the best systems from OpenAI and Anthropic. The launch posts lead with agents that remember context across weeks. The model card underneath is the more consequential document. Without its safety mitigations, Meta cannot rule out that the model would cross its own high-risk threshold for chemical and biological misuse.
Meta says that with mitigations in place the residual risk is "moderate or lower", the level it says supports its decision to deploy. A fuller safety report is promised "soon".
What Meta is selling
There are two models. Muse Wonder is available to developers through the Meta Model API and in Muse Code, Meta's coding product. Muse Wonder Full is reserved for Muse Code subscribers and, according to TechCrunch, is not served through the API at all. Both will later power Muse, Meta's consumer agent, starting with paying subscribers.
TechCrunch reports API pricing of $4 per million input tokens and $20 per million output tokens, roughly level with OpenAI's GPT-6 Sol and Anthropic's Opus 5.5. The model card lists a context window of one million tokens and input in text, images, video and audio, with text output. Meta says the models were trained at Prometheus, its gigawatt-scale data centre, and describes two headline traits it calls Active Memory and Active Focus. Active Memory carries relevant details across conversations. Active Focus keeps going at a hard task instead of giving up.
Reading the benchmark claims
Meta compares Muse Wonder with GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5, all at maximum reasoning effort, and says it leads on personal and professional agent tasks, memory and coding. The figures appear as charts in Meta's research post, and the model card explains how they were assembled. Where a third-party or provider-reported score exists, Meta uses it. Where none exists, it uses the highest of its own validated run, a self-reported score or a leaderboard figure.
The card does flag the soft spots. Competitor runs done in-house are "best-effort", with prompts and tools that "may not be optimally tuned for proprietary models". Several of the headline evaluations, including MemoryBench and FocusBench, are Meta's own. One coding comparison, SWE-Marathon, mixes different agent products and is "not a controlled common-harness comparison". None of that makes the results wrong. It does mean the independent picture will come from outside testers rather than launch charts.
The safety numbers behind the launch
The risk section is the most specific part of the card. Meta evaluates three catastrophic-risk domains under its Meta Superintelligence Scaling Framework. For cybersecurity and loss of control, it rates even the unmitigated model "moderate or lower". For chemical and biological risk it does not, and applies extra safety training and system-level classifiers before release.
On ExploitGym, Muse Wonder Full captured 92 flags. A verification pass found that only 21 used the intended vulnerability. The other 71 came through unintended ones. Because the benchmark's published ranking counts only intended exploits, Meta says 21 is the comparable figure. The model found many more ways in than the test was designed to measure. Meta also says it has kept "frontier cyber capabilities for defensive use" available.
Two biology tasks test help with designing and disguising DNA fragments. The card reports no score for either, because the model refused so often. It notes that refusals were dropped before scoring on those rows, unlike the capability benchmarks, where a refusal counts as a failure.
Where this leaves developers
For teams choosing a model, Muse Wonder is a credible new option at a familiar price. Readers can compare its cost with other models in our LLM cost calculator, a RECATOOLS tool. Meta's own guidance is to add guardrails such as human confirmation before irreversible actions, and it says the model is not intended for users under 18.
Still waiting on the full report
Muse Wonder follows Meta's first priced closed model, Muse Spark 1.1, launched in July. Its disclosure also compares well with Google's Gemini 4 Argon, whose security claims mostly came without numbers. Until Meta publishes the full safety report, the model card is the best evidence available, and it says more than the launch posts do.
Disclosure: RECATOOLS is written with the assistance of Claude, made by Anthropic, whose models Meta benchmarks Muse Wonder against. Readers should weigh that interest.
View author profile → · Editorial policy
