Meta's Muse Wonder Full Is Two Points Off the Top of the Index. You Cannot Buy It by the Token.

Six months after Meta closed the Llama line and three months after it started renting Muse Spark by the token, Meta Superintelligence Labs has shipped what it calls its first frontier models. Muse Wonder and Muse Wonder Full arrived today, October 12, 2026, "with state of the art performance for personal and professional agentic work," in Mark Zuckerberg's words on X.

The independent number backs the frontier claim. Artificial Analysis scored Muse Wonder Full at 56 on its Intelligence Index this afternoon, tied with Claude Sonnet 5.5 and two points behind Claude Opus 5.5 at 58, which makes Meta the third lab with a model in the top tier. It also took first place on AA's Harvey legal-agent benchmark and tied the top of AA's Coding Agents Index.

Then read the availability sentence: "Muse Wonder is available today for developers and builders through the Meta Model API and Muse Code while Muse Wonder Full is available today for Muse Code subscribers." The model that scores 56 is not on the API. The model on the API scores 52 and costs the same per token as the model that scores 58. We read the newsroom post, the 13-page model card, the API docs, the pricing and rate-limit pages, the Muse Code subscription terms, the geographic-use policy, and AA's analysis. Here is what held up.

Two Models, One Price List, One Subscription

The API price matches Opus 5.5 exactly. $4 and $20 is Claude Opus 5.5's list price, double Sonnet 5.5, Gemini 4 Argon, and GPT-6.1 Sol at $2/$10, and 60% below GPT-6 Astra and Claude Fable 5.1 at $10/$50. There is no batch discount, no long-context tier, and no published cache-write charge; web search is $2.50 per 1,000 queries. The rate limits are on Meta's pricing page, which a widely shared dev.to piece says do not exist.

The subscription trains on you by default. Meta's docs: "Muse Code subscriptions default to Contributor mode," and "In Contributor mode, you get more usage from your plan in exchange for letting Meta use your content to train future Meta models." You can switch it off. The API's Muse Wonder is Standard mode only, meaning no training on requests; zero data retention is a separate organization-level approval that disables stateful requests, background requests, file uploads, web search, and playground history.

"Max" on the API does not get you Wonder Full. Meta's docs say it in one sentence: "On Model API, 'max' on muse-wonder-1.0 doesn't select Muse Wonder Full, which is available only in Muse Code." The four Muse Code tiers are Everyday ($5), High ($15), Power ($50), and Maximum ($200); Wonder Full is on the top two.

Two dashes Meta chose. The model card gives no maximum output tokens and no knowledge cutoff, and neither does any Meta page we could find. Meta also publishes no parameter count; the only training detail is that both models were trained at the gigawatt-scale Prometheus data center in Ohio.

The Benchmarks, Sorted by Who Reported Them

House rules: independent numbers first, stated as fact with the source and index version; Meta's numbers labeled as Meta's, with the effort level, because Meta's table mixes them.

Artificial Analysis, Intelligence Index v4.3.2, October 12. Muse Wonder Full (max) 56, tied with Sonnet 5.5 (max) and Opus 5.5 at xhigh; Opus 5.5 (max) leads at 58. Muse Wonder (max) 52, alongside GPT-6.1 Sol. Wonder Full is in the top three on six of the ten index evaluations: first on CritPt (33%) and GDP.pdf (33%), third on SciCode (63%). On AA's Coding Agents Index, Wonder Full at max in Muse Code scores 68, level with Claude Code running Sonnet 5.5 (68) and above Opus 5.5 (66); Muse Wonder scores 63. On Harvey LAB-AA v1.1, the hallucination-gated legal-agent benchmark, Wonder Full is first at 13.1% and Wonder second at 10.0%, ahead of Grok 4.7 (9.4%), GPT-6 Astra (8.6%), GPT-6.1 Sol (6.9%), and Claude Fable 5.1 (6.4%).

The Wonder cost line deserves a second look. $4.52 per index task is in the same band as Opus 5.5 ($5.98) and Sonnet 5.5 ($5.46) for a lower score, and about 2.8 times what AA measures for Muse Spark 1.3. The speed and latency columns for Wonder Full print as dashes because nobody outside Muse Code can measure them.

Now the vendor table. These are Meta's self-reported numbers from the model card, max effort unless noted, against Claude Fable 5.1, GPT-6 Astra, and Claude Opus 5.5, all at max. Meta's claim of wins over all three "on coding, legal, and automation" holds on these rows:

Four honest observations:

• The same table shows the losses. Meta's card has Wonder Full trailing on terminal coding, on from-scratch programming (ProgramBench 59.0% for Full, 50.0% for Wonder), on health, on GDP.pdf, and on computer use. The newsroom post leads with the wins; the card prints both.
• AutomationBench appears twice with different numbers, and they are different metrics. Meta's row is a strict pass@1 of 45.4%. AA's own AutomationBench-AA inside the index scores Wonder Full at 71%, where it is third behind Gemini 4 Argon (77.5%) and Sonnet 5.5 (71.8%). Never put the two in one column.
• One row is at xhigh and the comparators ran in other harnesses. The SWE-Marathon 57.5% is Wonder Full at xhigh while every other Muse cell is max, and the Astra and Fable figures come from leaderboard runs in Codex and Claude Code respectively. The card says so on page four.
• "Active Memory" is a product description, not a mechanism. Meta defines it as "the ability to carry what matters to you across conversations and bring back the details that are relevant to the task at hand," measured on an internal MemoryBench with no architecture or storage disclosed. On Meta's own numbers the smaller Wonder scores higher on it than Wonder Full.

On terminal coding specifically, AA's Terminal-Bench 4.0 board has Sonnet 5.5 at 63.6% and Opus 5.5 at 59.6%; Wonder Full's 57% and Wonder's 52% are not top two.

Safety, Regions, and the Predecessor

The model card's chemical and biological line is unusually direct: "Without mitigations, we do not rule out the model reaching this threshold in the Chemical & Biological domain." With safety post-training and system-level classifiers, Meta rates all three risk domains "moderate or lower," and promises a fuller safety report it has not published.

The Meta Model API's Geographic Use Policy, updated today, designates both Wonder models as Standard Models, says the service is "not available in all jurisdictions," and lists 21 restricted territories, including China with Hong Kong and Macau, Russia, Iran, Pakistan, and Venezuela. There is no allow-list and no EU or UK exclusion.

Muse Spark 1.3, released September 2, stays on sale at $1.25 in / $0.15 cached / $4.25 out in Standard mode and $0.10 / $0.002 / $0.20 in Contributor mode, and scores 48 on AA's index. The Muse Code docs still name muse-spark-1.2 as the default model. Both Wonder models are closed weights with no self-hosting; Meta's open line is Muse Glimmer, a 30B Apache 2.0 model from August.

Where You Can Run It

Muse Wonder: the Meta Model API at api.meta.ai/v1, which speaks three protocols (Responses, Chat Completions, and Messages), as muse-wonder-1.0 in preview and Standard mode only, at 3,000 requests and 4 million tokens per minute per team; and Muse Code on every paid tier from the $5 Everyday plan up. Effort runs from minimal to max, with none rejected with an HTTP 400 and the default "model-determined." Prompt caching exists with a default lifetime of "evicted under pressure or after inactivity" and extended retention to 24 hours on request; there is no published minimum cacheable prompt and no separate cache-write price, although AA's cost breakdown books $1.92 of Wonder's $4.52 per task as cache writes. Zero data retention is available to qualified organizations by request through sales, and switching it on disables stateful requests, background requests, file uploads, web search, and playground history.

Muse Wonder Full: Muse Code Power ($50) and Maximum ($200) only, in both Standard and Contributor mode, with Contributor the default on every Muse Code plan.

Neither: the consumer Muse app, which as of today does not mention Wonder on its own page and promises Wonder "will roll out first to subscribers." Not Bedrock, Vertex, Foundry, or OCI. And nowhere in the 21 restricted territories.

The one practical note for teams already on Spark 1.3: the API price triples and the per-task cost on AA's index goes up about 2.8 times for a four-point gain on the index. Whether that trade is worth it depends on whether your work looks like Harvey LAB-AA or like Terminal-Bench.

Verified vs Unconfirmed: The Scorecard

Who Should Use It

Use Muse Wonder Full if you already live in Muse Code at the $50 or $200 tier and your work is legal, finance, or long-horizon coding agents, where it holds first places on independent boards. Switch Contributor mode off first if the code is yours.

Use the API's Muse Wonder only if you need its multimodal input (video and audio in a million-token window) at an Opus-class price; on intelligence per dollar it loses to Sonnet 5.5, Gemini 4 Argon, and GPT-6.1 Sol, all at half the per-token price and all scoring the same or higher.

Skip both if you need to pin a model version in production (the API model is a preview), serve the restricted territories, or know the output cap and cutoff before you commit. If what you want is Meta's frontier intelligence by the token, it is not for sale today.

The frontier is elsewhere: Claude Opus 5.5 (max) leads the v4.3.2 index at 58, with Sonnet 5.5 and Muse Wonder Full at 56 and GPT-6 Astra, Fable 5.1, and Gemini 4 Argon at 53. Meta is now in that conversation, on one benchmark board and in one subscription. Our model comparison has the current standings side by side.

Update, October 16, 2026

• October 12: Vals AI's Finance Agent v2 board, updated the same day, has Gemini 4 Argon first at 65.40% and Wonder Full fifth at 59.06%, behind Gemini 3.8 Flash and Muse Spark 1.2. Vals's own hosting of Harvey's Legal Agent Benchmark (task pass rate, a different harness from AA's) puts Muse Spark 1.2 first at 25.42% and Wonder Full fourth at 21.67%; the legal lead depends on whose harness you read.
• October 14: Vals AI's Vibe Code Bench 1-100 ranks Wonder Full first at 33.98%, ahead of Opus 5.5 at 30.36%, with Wonder sixth. A ComputeLeap piece republished on dev.to argued the launch "scores big and delivers little" because "a model launch is not a model. It is a product, an API, a trust relationship, and a competitive position"; it also misstated the rate limits. Meta published a health-research write-up built on Spark 1.3 and Wonder over five million de-identified health records, claiming up to tenfold researcher productivity, with no external review.
• October 15: Meta released Muse Spark Open, an 816B open-weight model that scores 39 on AA's index, completing a closed-top, open-below lineup.
• October 16: On AA's GDPval-AA v2.1 board Wonder Full sits second at 1857 Elo, between Opus 5.5 (1866) and Sonnet 5.5 (1838). Wonder Full still has no API, no price, and no cutoff.

Sources

• Meta Newsroom: Introducing Muse Wonder and Muse Wonder Full
• Meta: Muse Wonder model card
• Meta Model API docs: models
• Meta Model API docs: pricing and rate limits
• Meta: Muse Code plans
• Meta docs: Muse Code subscriptions and Contributor mode
• Meta Model API: Geographic Use Policy
• Artificial Analysis: Muse Wonder analysis
• Artificial Analysis: Muse Wonder Full model page
• Vals AI: Vibe Code Bench 1-100

*Not sure whether a frontier model you can only rent through one subscription fits how you work? Take the free 2-minute quiz and get matched. Then read the model Wonder Full ties on the index, which you can buy by the token at half the price: Claude Sonnet 5.5.*