Meta Muse Wonder: Frontier Model at $4/$20 per Million Tokens

Meta released Muse Wonder and Muse Wonder Full on October 12, 2026, its most capable models to date. On the Meta Model API, Muse Wonder costs $4.00 per million input tokens and $20.00 per million output tokens ($0.20 cached input) with a 1M-token context window, per Meta’s model page. Muse Wonder Full, the stronger variant, is not on the API at all: it ships only inside a Muse Code subscription.

The numbers on the wire

• $4.00 / $20.00 per million input / output tokens for Muse Wonder; $0.20 per million cached input tokens.
• 1M tokens of context; inputs are text, image, video and audio, output is text, per the model card.
• Personal Superintelligence Index, Meta’s internal composite of seven agent and knowledge-work evals: Wonder Full 59.1, Claude Opus 5.5 55.2, GPT-6 Astra 54.5, Muse Wonder 54.3, Claude Fable 5.1 49.9.
• DeepSWE 1.1 long-horizon coding (Datacurve third-party runs for the Muse models and Opus): Wonder Full 79.0, Wonder 78.8, Opus 5.5 74.2, GPT-6 Astra 74.1, Fable 5.1 67.4.
• SWE-Marathon ultra-long-horizon engineering: Wonder Full 57.5 (at xhigh effort) versus Opus 5.5 46.3, Fable 5.1 45.6 and GPT-6 Astra 42.5.
• Harvey LAB-AA legal agent work: Wonder Full 13.1, Wonder 10.0, GPT-6 Astra 8.6, Fable 5.1 6.4, Opus 5.5 4.2.
• MemoryBench, Meta’s internal long-conversation recall test: Wonder 85.6, Wonder Full 85.5, Opus 5.5 81.5, GPT-6 Astra 81.4, Fable 5.1 79.0.

Where it does not lead

Meta’s own scorecard shows Muse Wonder Full behind on several rows. Claude Opus 5.5 leads ProgramBench (70.5 versus 59.0), Terminal-Bench 4.0-AA (59.6 versus 57.1), HealthBench Pro (65.6 versus 62.3), OSWorld V2.1 partial credit (81.8 versus 78.4) the Artificial Analysis Intelligence Index (58 versus 56) and, narrowly, JobBench (75.4 versus 75.3). GPT-6 Astra leads GDP.pdf (34.2 versus 33.2). All comparisons are at max reasoning effort, and Meta notes that some rows, including SWE-Marathon, are not controlled common-harness comparisons because competitors ran in Codex or Claude Code.

How Meta built it

Meta says both models were trained at its gigawatt-scale Prometheus data center, 14 months after it began a ground-up overhaul of its AI efforts, and build on the Muse Spark line with larger base models. The research blog pitches two behaviors: Active Memory, which carries user context across conversations, and Active Focus, which keeps an agent working through blockers and checks in before important actions. Muse Wonder will also power the Muse agent, rolling out first to subscribers.

Safety receipts

Under its Meta Superintelligence Scaling Framework, Meta says it does not rule out the unmitigated Wonder Full model reaching the high-risk threshold in the chemical and biological domain, so it applies safety post-training and system-level classifiers there. It rates the unmitigated model “moderate or lower” for cybersecurity and loss of control. On cyber capability evals, Wonder Full scores 88.4% on CyberGym and 58.9% on ExploitBench. Meta says a full safety report is coming “soon.”

On the record

“Together, they mark the next step on our scaling ladder and a major leap toward personal superintelligence.” — Meta, Muse Wonder research blog

What to watch

The receipts to track: whether independent runs from Artificial Analysis and others reproduce Meta’s internal PSI and MemoryBench leads, whether Wonder Full ever reaches the API, and when the promised safety report lands. At $4 / $20, Muse Wonder is priced against the top of the frontier tier rather than under it, so adoption will show whether the coding and memory scores move developers. Meta’s newsroom post and TechCrunch carry the launch details.

FAQ

How much does Meta Muse Wonder cost?

$4.00 per million input tokens, $0.20 per million cached input tokens and $20.00 per million output tokens on the Meta Model API.

What is the difference between Muse Wonder and Muse Wonder Full?

Muse Wonder is available on the Meta Model API and Muse Code. Muse Wonder Full is Meta’s most advanced variant and is available only to Muse Code subscribers on the Power Usage and Maximum Usage plans.

How big is the Muse Wonder context window?

1 million tokens at inference, according to Meta’s model card.