Meta Releases Muse Wonder and Muse Wonder Full, Its First Frontier Models for Agents

Meta on October 12, 2026 released Muse Wonder and Muse Wonder Full, two models it says deliver state of the art results on personal and professional agent tasks. Muse Wonder is available through the Meta Model API and the Muse Code coding agent, while the stronger Muse Wonder Full is limited to Muse Code subscribers, the company said.

Meta says it began a ground up overhaul of its AI work 14 months ago; it introduced Muse Spark, the first model from the rebuilt Meta Superintelligence Labs, on April 8, 2026. TechCrunch calls Muse Wonder the company's first attempt at a frontier model, and it now competes with the top model releases from OpenAI, Anthropic and Google.

What did Meta release?

Meta released two models on October 12, 2026: Muse Wonder, sold through its API, and Muse Wonder Full, a stronger variant kept inside Muse Code. Both have a 1 million token context window and accept text, image, video and audio input, according to Meta's model card.

Meta pitches two features in its research blog. Active Memory carries what matters to a user across conversations; Active Focus keeps the model working on a hard task, trying other approaches and checking in before important actions. Meta AI chief Alexandr Wang told CNBC the model was codenamed Watermelon.

How does Muse Wonder compare with Claude and GPT-6 Astra?

In Meta's own benchmark table, Muse Wonder Full scores higher than Anthropic's Claude Opus 5.5 on 8 of 14 rows and lower on the other 6. Across all five models, Muse Wonder Full has the top score on 6 rows, Muse Wonder on 1, Claude Opus 5.5 on 6 and OpenAI's GPT-6 Astra on 1.

Two rows Muse Wonder Full leads, the Personal Superintelligence Index and MemoryBench, are Meta's internal tests. The model card says Meta's runs of rival models are best effort, and that SWE-Marathon is not a controlled comparison because GPT-6 Astra and Claude Fable 5.1 ran in Codex and Claude Code. Google's Gemini 4 Argon is missing, as The Verge points out, although Artificial Analysis lists it among similarly capable models.

What do independent tests show?

Artificial Analysis, a company that benchmarks AI models with its own tests, scores Muse Wonder Full at 56 and Muse Wonder at 52 on its Intelligence Index v4.3, up from 48 for Muse Spark 1.3, released six weeks earlier. Meta's own table lists Claude Opus 5.5 at 58 on that index, so the independent composite still ranks Anthropic's model ahead.

According to Artificial Analysis, Muse Wonder Full places in the top three on 6 of the 10 evaluations in the index, and its Terminal-Bench 4.0 score rose from 33% to 57%. On the firm's Coding Agents Index, Muse Wonder Full in Muse Code scores 68, level with Claude Code using Claude Sonnet 5.5 and above Claude Code using Opus 5.5 at 66. Factual recall is a weak spot: 30 on AA-Omniscience against 46 for Opus 5.5, offset by a hallucination rate that fell from 33% to 23%.

How much does Muse Wonder cost?

Muse Wonder costs $4 per million input tokens, $0.20 per million cached input tokens and $20 per million output tokens, according to Artificial Analysis and OpenRouter's listing. Muse Wonder Full has no token price and comes with the $50 or $200 monthly Muse Code plans (CNBC, Engadget).

Wang told CNBC the API would be priced "meaningfully below" leading models from OpenAI and Anthropic. Set against the price list Artificial Analysis published, that holds for the most expensive tier but not for Anthropic's Opus: Muse Wonder costs the same per token as Claude Opus 5.5, 60% less than GPT-6 Astra and Claude Fable 5.1 ($10 and $50), and twice as much as Claude Sonnet 5.5, Gemini 4 Argon and GPT-6.1 Sol.

Per Intelligence Index task, Muse Wonder costs $4.52, against $5.98 for Opus 5.5, $3.26 for GPT-6 Astra and $1.60 for Muse Spark 1.3, mostly because it reads about 7.69 million input tokens per task, Artificial Analysis says. Developers can also reach it through Vercel AI Gateway and OpenRouter, and Meta says it will soon power the Muse agent app, first for subscribers.

What safety limits did Meta set?

Meta's model card says it cannot rule out that Muse Wonder Full, without safeguards, reaches its high risk threshold for chemical and biological misuse, so it added safety training and system level classifiers. With those in place, Meta rates the remaining risk as "moderate or lower" in every domain it tests, and it says a full safety report "will be available soon."

What we don't know yet

• Whether Meta will publish weights for either model, and when Muse Wonder and Muse Code reach users in the European Union.
• When the full safety report will appear, and when Muse Wonder will power the Muse app; Meta has given no timeline for the app, The Verge notes.
• How Muse Wonder Full performs against Gemini 4 Argon on Meta's own benchmarks.