Muse Wonder

Muse Wonder is a series of multimodal reasoning models from Meta, released on October 12, 2026 for personal and professional AI agents. The series includes Muse Wonder and Muse Wonder Full, which Meta describes as the higher-capability variant. Meta launched the standard model through its developer API and Muse Code, while Full was restricted to eligible Muse Code subscriptions.^[1][2]

Models and access

Muse Wonder is in preview. The API model identifier is muse-wonder-1.0. Both variants are available in Muse Code, but choosing the max reasoning setting on the standard model does not select Full.^[3]

The Meta Model API accepts text, images, video, audio and PDF input for Muse Wonder and returns text. Its context window is 1,048,576 tokens. Supported capabilities include parallel tool calls, structured output, search grounding and computer use.^[3]

Meta's model page lists Power Usage at $50 per month and Maximum Usage at $200 per month. These are subscription prices, not a per-token API price for Full.^[2] Subscription benefits can vary by region. A subscription applies to the credential connected during Muse Code onboarding; additional API keys use pay-as-you-go billing. Reaching a plan's usage limit requires waiting for a reset or upgrading where a higher plan exists.^[4]

Pricing and data use

The October 2026 pay-as-you-go rates for Muse Wonder are:^[5]

Meta does not charge a higher token rate for long contexts. Web search grounding adds $2.50 per 1,000 search queries. The documented limits for muse-wonder-1.0 are 3,000 requests and 4,000,000 tokens per minute, shared across a team's API keys. Starting background responses also has a 600-per-minute team limit.^[5]

Standard mode does not use submitted content to train Meta models. Muse Wonder has no Contributor-mode pay-as-you-go option.^[3][5] Muse Code subscriptions default to Contributor mode, which permits Meta to use submitted prompts, code, files, context and responses for training in exchange for more usage. Subscribers can switch it off; future content is then excluded from training unless they enable it again. Both Wonder variants support either mode on subscription plans that include them.^[4]

Reasoning controls

Muse Wonder supports minimal, low, medium, high, xhigh and max reasoning effort. Setting none returns HTTP 400. Omitting the setting leaves the reasoning level to the model. Higher effort uses more reasoning tokens and generally increases latency and cost; max uses the same per-token rate as the other levels.^[6]

On Chat Completions, the setting is reasoning_effort; on Responses, it is reasoning.effort. Reasoning tokens count toward both billed output and the output-token limit, alongside visible text. A budget spent mostly on reasoning can therefore leave too little room for the answer.^[6]

For multi-step work, Responses can preserve reasoning context through previous_response_id or encrypted reasoning replay. Chat Completions does not expose readable reasoning content to external callers, so an external client cannot replay it through reasoning_content. Responses can return optional reasoning summaries, which are separate from encrypted replay data.^[6]

Active Memory and Active Focus

Meta calls the models' use of relevant earlier context Active Memory. The intended behavior is to retain important information across conversations and retrieve it when it becomes relevant to a new task. Active Focus is Meta's name for sustained work toward a goal: attempting alternatives, adapting to new information, asking for clarification and checking before important actions.^[1]

Meta's launch demonstrations included software construction and personalized planning. The company said Muse Wonder would subsequently power experiences in Muse, initially for subscribers, and artifact generation. That was a planned rollout, distinct from the developer access available on launch day.^[1]

Evaluation results

Meta's launch model card reported the following results, with both variants at max reasoning effort. Later leaderboards may use different configurations.^[7]

The Personal Superintelligence Index is Meta's internal seven-evaluation composite. MemoryBench is also internal and uses conversations of 64K-100K tokens. The card combines external results with internally generated comparisons; competitor runs are best-effort rather than uniformly tuned.^[7]

DeepSWE consists of 113 original engineering tasks across 91 repositories and five languages. Datacurve grades functional behavior with handwritten verifiers and uses mini-swe-agent for its published model comparisons.^[8]

SWE-Marathon measures larger projects with binary task completion: passing every verifier test earns full credit. Its v1.1 leaderboard, checked on October 13, 2026, listed Muse Wonder Full with Muse Code at 53.1% and Muse Wonder with Muse Code at 52.5%, both at max effort. The leaderboard runs eight trials per task and also compares other models through their own agent products.^[9]

The benchmark's original research distinguishes evaluation of a complete commercial agent from comparisons under a shared scaffold. Model, harness and benchmark version therefore belong with the score; a product-level result does not isolate the model's contribution.^[9][10]

Artificial Analysis separately reported an Intelligence Index score of 56 for Muse Wonder Full at max. Its v4.3.2 index aggregates ten independently run evaluations. This index is different from Meta's Personal Superintelligence Index.^[11]

Safety and limitations

Meta did not rule out its unmitigated Full model reaching its high-risk threshold in the Chemical & Biological domain. It assessed unmitigated Cybersecurity and Loss of Control risk as moderate or lower. With mitigations, Meta assessed all three domains as moderate or lower.^[7]

The model card warns of inaccurate or biased answers, multi-step reasoning failures and limitations in visual and spatial reasoning, especially 3D tasks. It recommends deployment-specific guardrails and human confirmation for irreversible agent actions.^[7]

References

1. ^1 ^2 ^3Meta AI. Muse Wonder: A Frontier Model For Personal & Professional AI Agents. October 12, 2026.
2. ^1 ^2Meta AI Developers. Muse Wonder. Accessed October 13, 2026.
3. ^1 ^2 ^3Meta Model API documentation. Models. Accessed October 13, 2026.
4. ^1 ^2Meta Model API documentation. Subscriptions. Accessed October 13, 2026.
5. ^1 ^2 ^3Meta Model API documentation. Pricing and rate limits. Accessed October 13, 2026.
6. ^1 ^2 ^3Meta Model API documentation. Reasoning. Accessed October 13, 2026.
7. ^1 ^2 ^3 ^4Meta AI. Muse Wonder Model Card. October 2026.
8. ^Datacurve. DeepSWE. Benchmark description and evaluation harness. Accessed October 13, 2026.
9. ^1 ^2Abundant AI. SWE-Marathon. Version 1.1 leaderboard and methodology. Accessed October 13, 2026.
10. ^Rishi Desai and colleagues. SWE-Marathon: Can Agents Autonomously Complete Ultra-Long-Horizon Software Work?. arXiv:2606.07682, June 5, 2026.
11. ^Artificial Analysis. Muse Wonder Full (Max). Intelligence Index v4.3.2. Accessed October 13, 2026.