Meta Launches Muse Wonder at $4 per Million Tokens, Safety Report Pending

Meta on Monday released Muse Wonder, a new model built for agent tasks, through its Meta Model API at $4 per million input tokens and $20 per million output tokens, with cached input discounted to $0.20 per million tokens. A stronger sibling, Muse Wonder Full, is available today for Muse Code subscribers.

The split matters for developers. Muse Wonder Full is not sold through the API; developers can use it only on Muse Code's Power Usage and Maximum Usage plans. Meta has not published API pricing for it.

The API model, using the model identifier muse-wonder-1.0, has a one-million-token context window. Web search through the API carries an additional charge of $2.50 per 1,000 search queries.

Independent testing puts the API model just short of the top tier. Artificial Analysis gives Muse Wonder a score of 52 on the Artificial Analysis Intelligence Index, which VentureBeat noted places it just one point behind GPT-6 Astra. Running that index cost $4.52 per task on average, so per-task costs may run higher than the token price alone suggests.

Meta's own benchmarks favor memory work. On MemoryBench, Muse Wonder scored 85.6 and Wonder Full 85.5, ahead of Claude Opus 5.5 at 81.5, GPT-6 Astra at 81.4 and Claude Fable 5.1 at 79.0. Those results are self-reported.

Meta says the models carry two new capabilities: Active Memory, which keeps relevant details across conversations, and Active Focus, which keeps working through difficult tasks, recovers from errors and checks in with the user before important actions.

On safety, Meta says it trained the models to avoid sharing private information without permission and to refuse harmful requests, and that it released them under its Meta Superintelligence Scaling Framework. It said a full report with additional safety evaluations will be available soon, but gave no date.

The timing stands out. Four days earlier, the UK's Information Commissioner's Office said it was making enquiries with Meta, OpenAI and Anthropic about how they test and deploy AI agents, as AI Briefing News reported.

Meta did not say when the safety report will be published, whether Muse Wonder Full will reach the API, or whether it will release model weights for either model. Its announcement made no mention of open weights.

Muse Wonder is Meta's most capable model so far and will soon power experiences in Muse, and will roll out first to subscribers. Meta says it began an overhaul of our AI efforts 14 months ago, and the new models build on Muse Spark, the first model from Meta Superintelligence Labs, released in April.