Meta Muse Wonder and Muse Wonder Full, Meta's AI agents, have been released for personal and work tasks. Both models lead with Active Memory, which remembers context across conversations, and Active Focus, which keeps hard tasks moving. Muse Wonder is available through the Meta Model API and Muse Code, while Wonder Full is available on select Muse Code subscription plans.
The models, announced on October 12, 2026 (October 13 in Korea), are the result of a 14-month effort by Meta to rebuild its AI stack. Meta trained both on its Prometheus data center, and its research announcement described the training infrastructure as "gigawatt-scale."
Source: X @AIatMeta
An Agent That Remembers and Stays on Goal
Active Memory remembers important user context across multiple conversations and retrieves what the current task needs. Active Focus refers to the ability to try a different approach when work stalls, take in new information, and ask questions or request confirmation before important actions when needed. Meta says it evaluated these with MemoryBench, FocusBench, and long-running software tasks.
Source: X @AIatMeta
Performance Compared Across 14 Evaluations
The Meta Model API page publishes 14 evaluations covering agentic work, coding, and general ability. On the Personal Superintelligence Index in Meta's chart (an internal composite for personal and knowledge work), Muse Wonder Full scored 59.1 and Claude Opus 5.5 scored 55.2. On DeepSWE 1.1, Muse Wonder Full scored 79.0 and Muse Wonder scored 78.8 (at maximum reasoning settings). On SWE-Marathon, Wonder Full scored 57.5 and Wonder scored 53.1 at the xhigh setting. In Meta's published chart, Wonder Full beat every compared model on 6 of the 14 evaluations. These are results on individual evaluations and do not mean overall superiority. According to the model card, the external results and execution harnesses vary by evaluation. Some comparison models on SWE-Marathon are Codex or Claude Code harness results, so they cannot be treated as identical conditions.
Source: X @AIatMeta
Developers can look at the conditions of other models alongside these on the AI model performance and pricing comparison page.
Where and How to Use the API and Muse Code
Muse Wonder is available through the Meta Model API and Muse Code. The context window shown on the API model page is 1,048,576 tokens. Muse Wonder pricing on standard pay-as-you-go is $4.00 per 1M input tokens, $0.20 per 1M cached input tokens, and $20.00 per 1M output tokens. To use Muse Wonder, create a Meta Model API account and issue an API key from the model page; in Muse Code, log in to the CLI to get started. You can start choosing a model and issuing an API key on Meta's Muse Wonder developer page.
Source: X @MetaforDevs
Muse Wonder Full has no published API token pricing. Muse Code subscriptions cost $50 per month for Power Usage and $200 per month for Maximum Usage. Meta's help page says Power Usage provides 20 times the usage of Everyday Usage and Maximum Usage provides 80 times, and that Maximum includes the benefits of Power Usage. Subscriptions apply only to the Muse Code CLI, and any additional API keys are billed pay-as-you-go. Plan availability may vary by region. Meta said it plans to roll this out to the general Muse app in stages, starting with subscribers. You can check the details in Meta's Muse Code subscription guide.
Source: X @MetaforDevs
If you are using Meta Muse for the first time in Korea, you can also refer to the sign-up and installation guide for Korea.
Safety Evaluations Meta Published
In the model card, Meta said it evaluated Wonder Full, the most capable variant. It reported BioTIER refusal rates of 98.7% for English and multilingual prompts each and 98.2% for adversarial variants. The cyber jailbreak attack success rate was 0.8% for both single-turn and multi-turn CyberBench, and 9.0% and 8.0% on RedCode. The model card adds that of the 92 flags in ExploitGym, 21 were verified as exploiting the intended vulnerabilities. These are results from evaluations and conditions set by Meta.
Source: X @AIatMeta
Infrastructure Scale-Up Discussed in the Interview
In episode 35 of the public interview series Cheeky Pint, Mark Zuckerberg and Alexandr Wang talked about Muse Wonder and the future of personal agents. The episode description covers the context of Meta's infrastructure expanding to more than 5GW. That figure does not refer to the power of any single facility used to train the models. Separately, the research announcement described the Prometheus training infrastructure as "gigawatt-scale."
Source: X @MetaNewsroom
You can also read about Zuckerberg's vision for personal AI agents in this related post.