Updated on October 13, 2026
Muse Wonder is Meta's frontier AI model, purpose-built for autonomous agents that handle complex legal, coding and professional tasks over extended sessions. Announced on October 12, 2026, the series ships in two variants, Muse Wonder and Muse Wonder Full, both with a context window of roughly one million tokens. The API costs $4 input and $20 output per million tokens, while Wonder Full is locked behind Muse Code subscriptions starting at $50 a month. Neither model comes with downloadable weights, a sharp break from Meta's earlier Llama strategy.
Muse Wonder takes at least text and images as input, judging by Meta's official demos, and returns text. Two headline features set it apart from Muse Spark, the lighter model released in April 2026. Active Memory retains key facts from one conversation to the next, while Active Focus keeps the model pushing through roadblocks rather than stalling and checking back in with you at every turn.
Meta's own demos show the model planning a Saturday run from personal preferences, understanding a real estate lease, and building a playable 3D kitchen game from a handful of reference images. All of this happens inside a browser, with the model chaining steps on its own.
On Harvey LAB-AA, a legal agent benchmark, Wonder Full scores 13.1 against 8.6 for GPT-6 Astra and 4.2 for Claude Opus 5.5. SWE-Marathon, which measures very long coding projects, puts Wonder Full at 57.5 versus 46.3 and 42.5. DeepSWE 1.1 tells a similar story, with Wonder Full at 79.0 and Wonder at 78.8, both above the roughly 74 posted by the two rivals.
The independent Artificial Analysis Intelligence Index paints a different picture. Wonder Full lands at 56 and Wonder at 52, while Claude Opus 5.5 leads at 58. The series' real strength is sustained autonomous agent work, not general-purpose benchmarks.
| Benchmark | Wonder Full | Wonder | GPT-6 Astra |
|---|---|---|---|
| Harvey LAB-AA (legal) | 13.1 | 10.0 | 8.6 |
| SWE-Marathon (long code) | 57.5 | 53.1 | 42.5 |
| DeepSWE 1.1 | 79.0 | 78.8 | 74.1 |
| Artificial Analysis Index | 56 | 52 | 53 |
Muse Wonder (model ID muse-wonder-1.0) is live on the Meta Model API. Cached input drops to $0.20 per million tokens, and built-in web search costs $2.50 per 1,000 queries. Rate limits sit at 3,000 requests and 4 million tokens per minute.
Wonder Full has no per-token price. Access requires a Muse Code subscription, either Power Usage at $50 a month (20 times the base quota) or Maximum Usage at $200 a month (80 times). Meta Platforms, which also runs Instagram and WhatsApp, has not announced whether token-based pricing for Wonder Full will follow. Prices may change, so always check the official site for the latest figures.
Muse Wonder is Meta's frontier-class model for long-running agent tasks, whereas Muse Spark is the lightweight, cheaper model launched in April 2026. Both coexist in Meta's lineup, but Wonder targets sustained legal, financial and coding work that Spark was never designed for. Think of Spark as the quick assistant and Wonder as the one that can spend hours on a single project.
No. Neither Muse Wonder nor Muse Wonder Full ships with downloadable weights, and self-hosting is not an option. This marks a clear departure from Meta's Llama line, where weights were publicly available. Access goes through the Meta Model API for Wonder, or a Muse Code subscription for Wonder Full.
Developers can call Muse Wonder (muse-wonder-1.0) via the Meta Model API at $4 per million input tokens and $20 output. Wonder Full requires a Muse Code Power plan at $50 a month or Maximum at $200 a month. The consumer Muse app will get Wonder soon, subscribers first, with no date confirmed.
It depends on the task. Wonder Full leads Harvey LAB-AA (13.1 vs. 4.2 for Opus 5.5), SWE-Marathon and DeepSWE 1.1 by wide margins. However, Claude Opus 5.5 still tops the Artificial Analysis Intelligence Index at 58 versus 56 for Full and 52 for Wonder, and it wins Terminal-Bench 4.0 and HealthBench Pro. For long-horizon code and legal agents, Wonder has the edge; for general reasoning, the race is tight.
Verdict: Long-running coding sprints and legal document marathons are where this model earns its keep. Development teams that need an agent capable of sustaining focus across hundreds of steps, at an API price matching Claude Opus 5.5, will find Muse Wonder worth testing, keeping in mind that the strongest variant lives behind a Muse Code subscription.
« An all-in-one creative platform that brings together more than 30 image and video models in a single interface. It includes camera controls that let you direct like a real filmmaker, characters that remain consistent from scene to scene, and ready-to-use tools for ads and short-form content »
Claude Opus 5.5 
GPT-6 Astra 
Claude Fable 5.1 
