Meta priced Muse Wonder into the frontier bake-off

Meta spent most of 2026 talking about personal agents and a light model called Muse Spark. On 12 October the company published something harder to wave away: a public price for a model it wants compared with Opus and Astra.

Wonder is $4/$20 on the API; Full stays behind Muse Code

Meta Superintelligence Labs released Muse Wonder and Muse Wonder Full. The model card dates the ship to 12 October 2026. Context is 1M tokens. Inputs include text, image, video, and audio; output is text. Meta’s model page lists the API id as muse-wonder-1.0.

Wonder is the SKU you can call. Josh Walters’ developer post puts it on Meta Model API and Muse Code at $4 per million input tokens, $0.20 cached, and $20 output. That matches Opus 5.5 list rates and sits far above Spark 1.3’s earlier $1.25/$4.25, as Artificial Analysis notes.

Full scores higher on Meta’s own sheet, and Meta kept it off the public API. Access runs through Muse Code’s Power plan ($50 a month) and the new Maximum plan ($200 a month, described as 80x Everyday usage). The research blog’s language about bringing frontier capability to everyone sits awkwardly next to that gate.

Wonder also has further to travel before it is the Muse consumer agent. Meta says the model will “soon” power Muse, starting with subscribers. Robert Hart at The Verge flagged the missing calendar. Until Meta names a day, that line is roadmap copy.

Agent API teams and Muse Code Power buyers feel it first

Teams already wiring agents through an OpenAI-compatible client who can absorb $4/$20. Muse Code subscribers on Power or Maximum who run long coding loops. Evaluation desks lining Wonder up against Opus 5.5, GPT-6 Astra, Claude Fable 5.1, and Gemini 4 Argon on multi-hour work.

Free Muse users leave these posts without a Wonder runtime notice. Everyday Muse Code without Power or Maximum never reaches Full. A procurement line for “Meta AI in chat” still does not purchase muse-wonder-1.0.

Artificial Analysis shows a six-week climb; Meta's chart still splits

Artificial Analysis, on its own Intelligence Index v4.3, scores Wonder Full at 56 and Wonder at 52 at max effort. Six weeks earlier, Muse Spark 1.3 sat at 48. AA also reports fewer output tokens than Spark and about $4.52 per Index task for Wonder at max, against roughly $2 for Argon at published high effort. Capability rose. Unit cost rose with it.

Meta’s launch table is narrower than the quotes around it. Wonder Full leads DeepSWE v1.1 (79.0), SWE-Marathon (57.5 at xhigh), Harvey LAB-AA (13.1), AutomationBench-AA (45.4), and Meta’s internal Personal Superintelligence Index (59.1). Standard Wonder takes MemoryBench at 85.6. Those rows match the pitch Meta wants: long jobs, recall, professional agents.

On the same Meta chart, Claude Opus 5.5 still leads Terminal-Bench 4.0-AA, ProgramBench, OSWorld V2.1 (partial), HealthBench Pro, JobBench, and the Artificial Analysis Index row (58 against Full’s 56). Meta chose to publish a composite where Opus finishes ahead.

“Active Memory” and “Active Focus” are Meta labels for training targets. TechCrunch carried Meta’s MemoryBench pride; the cleaner independent signal is AA’s Index move beside that split board. Long-horizon coding and memory look stronger. Terminal work, clean-room programming, and desktop computer use still favour other stacks on Meta’s numbers.

Muse dates and the full safety report are still missing

Wonder’s debut inside Muse apps has no public ship date. The full safety report is still promised rather than posted. The model card says that without mitigations Meta does not rule out a high Chemical & Biological risk threshold under its Superintelligence Scaling Framework, and that with mitigations it rates residual risk as moderate or lower. Outside auditors have not signed that grade in a published report.

Eval harnesses disagree across vendors. Meta’s model card notes SWE-Marathon uses different agent products for some rivals, including Codex and Claude Code. A shared harness could reorder Terminal-Bench and ProgramBench. These posts also leave Wonder versus Argon on your repository unsettled. AA still marks knowledge as a relative soft spot even while hallucination rates improve.

Zuckerberg’s “personal superintelligence” line remains a slogan until someone defines a test you can fail.

Run one marathon before you re-rank the stack

API teams: put muse-wonder-1.0 on a single multi-hour agent this week, price the trace at $4/$20 with cache, and keep Opus or Argon in the failover path wherever terminal or clean-room coding still decides the job.

Muse Code Power or Maximum subscribers: try Full on DeepSWE-class work. MemoryBench alone is a thin reason to buy Maximum.

Consumer Muse users can leave connectors alone until Meta names the runtime in product.

Meta now has a priced model that belongs in the same bake-off as Opus and Astra, with a clearer long-horizon story than Spark offered. Ranking the field still takes a task, a harness, and a bill. Start there.