Meta released Muse Wonder and Muse Wonder Full Monday, the models TechCrunch described as the company's first attempt at a frontier model, and Meta's own documentation says it could not rule out that they reach its "high risk" threshold for chemical and biological misuse before safeguards were added. Meta says added protections bring the risk down to an acceptable level. But a fuller safety report is still pending, so businesses adopting the models must, for now, rely largely on the company's summary judgment.

What Happened

The model card, the technical document that accompanies an AI release, sets out the assessment under Meta's internal rulebook, the Meta Superintelligence Scaling Framework. Under that framework, a model reaches the high-risk threshold when deploying it could substantially contribute to a threat scenario Meta has judged severe. Without mitigations, Meta writes, it does not rule out Muse Wonder reaching that threshold in the chemical and biological domain. It therefore applied safety training and system-level classifiers, filters that screen requests and responses, aimed at that area.

Meta rated the unprotected model "moderate or lower" for cybersecurity and for loss of control, the risk that humans can no longer direct or shut down an AI system. With mitigations in place, it reached the same rating in all three areas, which it says supports the decision to deploy. The tests were run on Wonder Full, the most capable version, and Meta says they cover the whole family. The card says the work included red-teaming, threat modeling with outside experts and input from government bodies, and that an internal governance process reviewed the result before release.

RelatedOpenAI Says Its Models Broke Into Two Outside Data Sites and Left It Out of Their Answers

Meta's tests point to strong cyber skills. Wonder Full scored 88.4% on the CyberGym benchmark and 58.9% on ExploitBench, which measures exploit writing. On a chemistry refusal test Meta built itself, the model refused 100% of high-risk English prompts, and it refused 98.7% on SecureBio's BioTIER biology benchmark.

Meta's announcement and research blog say a full report with additional safety evaluations "will be available soon," without giving a date. The blog says the models carry Meta's strongest safeguards yet against chemical, biological, radiological, nuclear and cyber misuse, while keeping advanced cyber capabilities available for defensive work such as finding flaws in code.

Muse Wonder is available through the Meta Model API at $4 per million input tokens and $20 per million output tokens, TechCrunch reported. Wonder Full is limited to subscribers of Meta's Muse Code product. TechCrunch reported that Meta says the models match Fable 5.1 and GPT-6 Astra on benchmarks; those figures have not been independently verified. Both models will power Meta's Muse agent, and they were trained at Meta's Prometheus data center in Ohio.

Why It Matters

For the chemical and biological risk, Meta's safety case rests partly on system-level filters the company operates, not only on how the model itself was trained. Companies that build on the API are relying on those protections staying in place and working, and on an assessment no outside party has yet published.

The card puts responsibility on deployers as well. Meta tells developers to evaluate whether the model suits their use and to add guardrails, such as human confirmation before irreversible actions, when it can act in the real world. Help Net Security highlighted that recommendation. The card also lists limits, including errors in multi-step reasoning and weaker visual and spatial reasoning. That matters because Meta pitches the models for agent work on insurance claims, leases and app building, and says they are trained to avoid sharing private information without permission.

For boards and risk officers, the practical step is to treat Meta's ratings as a starting point. Ask vendors for the full evaluation before moving sensitive workflows onto these agents.

What to Watch

The key signal is the release of Meta's full safety report and how much detail it gives on the chemical and biological tests. Also watch whether outside evaluators, including government testing bodies, publish their own results. Muse Wonder's planned rollout in Meta's Muse app, starting with subscribers, will widen exposure beyond developers.