Meta Muse Wonder and Muse Wonder Full: price, availability and what Artificial Analysis found

On October 12, 2026, Meta Superintelligence Labs released Muse Wonder and Muse Wonder Full, its first models aimed squarely at the top tier of frontier models. Meta calls them "our most capable models to date" and says they're built for "personal and professional agentic tasks." Muse Wonder is on Meta's paid API at $4 per million input tokens and $20 per million output tokens. Muse Wonder Full, the stronger of the two, isn't on the API at all: you get it through a paid Muse Code subscription.

This story is based on Meta's newsroom post, research blog, developer site and model card, and on Artificial Analysis's evaluation pages, all checked October 14, 2026, plus reporting from TechCrunch and VentureBeat.

The short version

• What: two closed-weight models from Meta Superintelligence Labs, built on the Muse Spark approach and trained at Meta's Prometheus datacenter.
• Price: Muse Wonder (muse-wonder-1.0) costs $4 input, $0.20 cached input and $20 output per million tokens.
• Wonder Full: no API. It comes with Muse Code's Power ($50/month) and Maximum ($200/month) plans.
• Independent scores: Artificial Analysis gave Wonder Full 56 and Wonder 52 on its Intelligence Index v4.3, up from 48 for Muse Spark 1.3.
• Muse app: Meta says Wonder will "soon" power more of the Muse agent, starting with subscribers.

What are Muse Wonder and Muse Wonder Full?

They're a step up from Muse Spark, the model family Meta has shipped since April and that we covered in our Muse Spark explainer. TechCrunch described Spark as the lighter, lower-latency line and Wonder as larger and more capable. Meta hasn't published parameter counts or architecture details.

Meta highlights two behaviors:

• Active Memory. Meta says the models carry what they've learned about you across conversations and pull back details that matter to the current task. TechCrunch reported that Meta cited an internal test called MemoryBench, on which Meta says Wonder beat Anthropic's and OpenAI's top models. The model card describes MemoryBench as a Meta-run test of recalling facts across 64,000- to 100,000-token conversations, but it doesn't report scores. Treat the memory claim as vendor-reported and unchecked.
• Active Focus. Meta's term for sticking with a hard goal: trying other approaches, getting past blockers, asking for clarification, and checking in "before taking important actions."

Like Muse Spark, Wonder is closed. Meta's model card says use is governed by its Model API terms, and it says nothing about open weights. At publication, Meta's only open-weight Muse model was Muse Glimmer; on October 15, Meta released the 816B-parameter Muse Spark Open.

How much does Muse Wonder cost, and where can you use it?

Last verified: October 14, 2026 (Meta for Developers model page)

That puts Muse Wonder at the same list price as Claude Opus 5.5 ($4/$20; see which Claude model to use) and well above Muse Spark 1.3 ($1.25/$4.25). Meta's developer site says Power includes 20 times the usage of its Everyday plan and Maximum 80 times. These are Muse Code plans; don't confuse them with the similarly named $20 and $100 plans for the Muse consumer app.

List prices don't tell you the bill. Artificial Analysis estimated it cost $4.52 on average to run one of its Intelligence Index tasks on Muse Wonder (max), against $5.98 for Opus 5.5 and $3.26 for GPT-6 Astra. Most of that was input: Wonder read 7.69 million input tokens per task, mostly cached. Our piece on why token pricing isn't the whole story explains why.

One privacy note: VentureBeat reported that Muse Code subscriptions default to Meta's "Contributor" mode, which lets Meta train on your prompts and code unless you switch it off. The pay-as-you-go API doesn't.

On the consumer side, Meta's newsroom post says Wonder "will soon power more experiences in Muse," Meta's personal agent app, "rolling out first to subscribers." TechCrunch wrote that both models will power Muse. Meta hasn't said when, or whether Wonder Full will reach the app.

What did Artificial Analysis find?

Artificial Analysis, an independent evaluator, published its results on October 12. Its Intelligence Index v4.3 combines ten evaluations of reasoning, knowledge, coding and agentic work into one score. Both Muse models were run at max reasoning effort.

Last verified: October 14, 2026 (Artificial Analysis leaderboard, index v4.3.2, and its October 12 Muse Wonder write-up)

So Wonder Full is tied for the second-highest score on the board, behind only Opus 5.5 at max effort. Its leaderboard row sits fourth because two Anthropic configurations share the score. The full picture is more mixed:

• Where Wonder Full led: narrow first places on CritPt and GDP.pdf (33% each, with the nearest rivals at 31–32%), and second on GDPval-AA v2.1 (1857 Elo), within the confidence intervals of Opus 5.5 (1866) and Sonnet 5.5 (1838).
• Biggest jump: Terminal-Bench 4.0 rose from 33% for Spark 1.3 to 57%.
• Weak spot: knowledge. On AA-Omniscience, Wonder Full scored 30 against Opus 5.5's 46, though its hallucination rate fell from 33% to 23%.

Limitations. One composite number hides big per-test swings, and a 2-point gap can be noise. Wonder Full has no public API, so outside developers can't reproduce its runs the usual way. These are one evaluator's results at one effort setting.

Meta's own launch charts against GPT-6 Astra, Claude Fable 5.1 and Opus 5.5 are vendor-reported, and its model card says some rivals ran in different harnesses, such as Codex or Claude Code.

What does the safety report say?

Meta's model card says that, before mitigations, it couldn't rule out Wonder Full reaching its "high-risk" threshold for chemical and biological weapons, so it added safety training and classifiers. With those in place, Meta rated it "moderate or lower" in every domain under its Superintelligence Scaling Framework. Meta's newsroom post says a fuller safety report is coming "soon." Our guide on how to read a system card covers how to weigh claims like these.

Meta also recommends human confirmation before irreversible agent actions. For an agent that reads your email and the web, that matters; see our prompt injection explainer.

What's confirmed, reported and uncertain

• Confirmed (Meta): the October 12 release; model IDs; 1M context and input types; API prices; Wonder Full limited to Muse Code's Power and Maximum plans; closed weights; Scaling Framework conclusions.
• Confirmed (Artificial Analysis): Intelligence Index scores of 56 and 52 at max effort, per-evaluation results, cost per task and speed.
• Vendor-reported: "state-of-the-art" agentic performance, MemoryBench results and Meta's comparison charts.
• Reported: Contributor mode as the Muse Code default (VentureBeat); both models powering Muse (TechCrunch).
• Uncertain: parameter counts; when Wonder reaches the Muse app; whether Wonder Full will get an API; the full safety report's contents.