Meta Muse Wonder: The Benchmarks, the Price, and Where Claude Still Wins

Meta Muse Wonder launched on October 12, 2026 as Meta’s new family of frontier models. The larger version, Muse Wonder Full, wins 7 of the 14 rows on Meta’s own benchmark chart against Claude Fable 5.1, GPT-6 Astra and Claude Opus 5.5. It leads hardest on long coding projects and legal agent work. Artificial Analysis, which tests models independently, scores it 56, level with Claude Sonnet 5.5 and two points behind Claude Opus 5.5.

The catch is in the details Meta’s launch post skips. The version you can buy through the API is the smaller Muse Wonder. It costs $4 and $20 per million tokens, exactly what Claude Opus 5.5 costs, and on Meta’s own chart it beats Opus 5.5 on only 5 of 14 rows. This article covers what Muse Wonder is, the full benchmark table, the independent check, the price, how to use it today and what Meta’s safety card discloses.

Mark Zuckerberg announced the launch on X with Meta’s full scorecard attached:

The Key Takeaways

• Two models, two doors. Muse Wonder is on the Meta Model API and Muse Code. Muse Wonder Full is only in Muse Code, on the $50 and $200 a month plans.
• Wonder Full wins 7 of 14 rows on Meta’s chart, led by SWE-Marathon (57.5 vs Opus 5.5’s 46.3) and DeepSWE 1.1 (79.0 vs 74.2). Two of those seven wins are on Meta’s own internal benchmarks.
• Claude Opus 5.5 still tops the independent index. Artificial Analysis Intelligence Index v4.3.2 puts Opus 5.5 at 58, Wonder Full at 56 and Wonder at 52.
• Same price as Opus 5.5. Muse Wonder costs $4 input, $0.20 cached and $20 output per million tokens, identical to Opus 5.5, although Alexandr Wang said it would be priced “meaningfully below” OpenAI and Anthropic.
• No open weights. Neither the launch post nor the model card mentions releasing them, and the Muse Spark weights promised in September have not shipped either.

What Is Meta Muse Wonder?

Meta Muse Wonder is a family of two frontier AI models from Meta Superintelligence Labs, released on October 12, 2026 and built for agents that work on your behalf for hours. Zuckerberg described the split in his launch thread: “Muse Wonder for fast frontier performance and Muse Wonder Full for the full version of the model.”

Muse Wonder is the successor class to Muse Spark, the family Meta launched in April 2026 and last updated with Muse Spark 1.3 in September. Zuckerberg called Wonder “a much larger and more advanced class of models that brings us to the frontier.” Meta says it was trained at its gigawatt-scale Prometheus data center, fourteen months after the company began what it calls “a ground-up overhaul” of its AI work. CNBC reports the internal codename was Watermelon, which matches the watermelon emoji Zuckerberg posted as a teaser on September 2.

The specs on Meta’s Muse Wonder model card are short. Both models take a 1 million token context window. They accept text, images, video and audio as input and answer in text only. Meta’s official account confirmed the availability split the same afternoon:

Active Memory

Active Memory is Meta’s name for carrying context about you across conversations and pulling up the right detail when it becomes relevant. Meta calls reliable memory one of the most critical steps toward useful personal agents. Its proof is MemoryBench, an internal test of recall across conversations of 64,000 to 100,000 tokens with deliberately conflicting and updated facts. Both Muse Wonder models top that row, at 85.5 and 85.6.

Active Focus

Active Focus is Meta’s name for staying on a hard task: trying another approach, working through blockers and checking in before important actions. Meta AI chief Alexandr Wang told CNBC the goal was an agent that does not stall when a dinner booking hits a snag. “We specifically tried to train the model to be able to try other approaches instead of stalling and then checking back in with the user at any time,” he said. Meta measures this with FocusBench, another internal test, and with SWE-Marathon, a third-party benchmark of very long software projects.

The Full Meta Muse Wonder Benchmark Table

Meta’s scorecard compares both Muse Wonder models with Claude Fable 5.1, GPT-6 Astra and Claude Opus 5.5, all at max reasoning effort. The table below is transcribed from the chart Meta published, with the best score in each row in bold. Every number is reported by Meta.

Where Muse Wonder Full wins

Muse Wonder Full wins long-horizon coding and legal work by wide margins. On SWE-Marathon it scores 57.5 against 46.3 for Opus 5.5 and 42.5 for GPT-6 Astra. On Harvey’s legal agent benchmark it more than triples Opus 5.5, 13.1 to 4.2, though every model on that row completes only a small share of tasks. It also takes DeepSWE 1.1, AutomationBench-AA and, narrowly, Vals Finance Agent V2. The Muse Wonder family takes 7 rows in total: six for Full and MemoryBench for the smaller model.

Where Claude Opus still wins

Claude Opus 5.5 wins 6 rows on Meta’s own chart: HealthBench Pro, JobBench, OSWorld, Terminal-Bench 4.0, ProgramBench and the Artificial Analysis Index. The ProgramBench gap is the largest on the page, 70.5 to 59.0. Remove the two rows Meta built itself, the Personal Superintelligence Index and MemoryBench, and Opus 5.5 wins more of the remaining twelve rows than Muse Wonder Full does, six to five. GPT-6 Astra takes the last row, GDP.pdf.

How Meta picked the numbers

Meta’s model card discloses three choices that favour its own column. First, where no third-party score exists, Meta uses “the highest of our internally validated run, self-reported score, or official leaderboard score.” Second, Meta calls its own runs of rival models “best-effort” and says prompts and tools “may not be optimally tuned for proprietary models.” Third, the SWE-Marathon headline uses Wonder Full at xhigh effort while the rest of the table runs at max, and the GPT-6 Astra and Fable 5.1 scores on that row come from Codex and Claude Code, so Meta itself says the row is “not a controlled common-harness comparison.” None of this makes the numbers wrong. It does mean the margins are a ceiling, not a floor. If these benchmark names are new to you, our guide to what AI benchmarks actually measure covers the main ones.

What Independent Testing Says About Muse Wonder

Independent testing agrees with Meta’s headline and with its limits: Muse Wonder Full is a frontier model, but not the top one. Artificial Analysis, which runs its ten-evaluation Intelligence Index itself, scores the models as follows on version 4.3.2.

Muse Wonder Full scores 56, two points behind Opus 5.5. Claude Sonnet 5.5, which Meta’s chart leaves out, also rounds to 56 and edges Wonder Full on the unrounded figure, 56.0 to 55.8. That makes Wonder Full the third-highest model on the index, three points ahead of Claude Fable 5.1 and GPT-6 Astra. Muse Wonder, the version on the public API, scores 52. That puts it one point below Fable 5.1 and Astra and six below Opus 5.5. The scores Meta printed in its own Artificial Analysis row match the independent figures exactly. Our monthly ranking of the best AI models tracks how these positions move.

Early hands-on reaction is warmer than the index. Shopify chief executive Tobi Lütke posted his verdict the same hour. He is not a neutral tester: Lütke sits on Meta’s Advisory Group, Muse has had the entire Shopify catalogue built in since launch, and Meta announced Shop Pay payments for Muse at Connect. Treat it as one experienced user’s impression, not a measurement:

Meta Muse Wonder Pricing: The Same Price as Claude Opus

Muse Wonder costs $4 per million input tokens, $0.20 cached and $20 per million output tokens on the Meta Model API. That is identical to Claude Opus 5.5 on every line. Alexandr Wang told CNBC the model would be priced “meaningfully below” the leading models from OpenAI and Anthropic. That holds against Claude Fable 5.1 and GPT-6 Astra, and it does not hold against Opus 5.5.

What the price buys you

At the same per-token price, Claude Opus 5.5 scores six points higher on the independent index than Muse Wonder and wins 9 of the 14 rows on Meta’s own chart against it. Muse Wonder beats Opus 5.5 on five: Harvey LAB-AA, GDP.pdf (by 0.2 points), MemoryBench, DeepSWE 1.1 and SWE-Marathon. If your workload is long coding projects or legal document work, those rows are the reason to test it. Claude Sonnet 5.5 is the harder comparison: at $2 and $10, half Muse Wonder’s price, it scores 56 on the same index, four points above Muse Wonder. For anything else, the price does not undercut Anthropic. Muse Wonder also costs 3.2 times as much as Muse Spark 1.3 for input and 4.7 times as much for output. Muse Wonder Full has no per-token price at all, because it is not on the API.

Muse Code and Muse app plans

Muse Code, Meta’s terminal coding agent, has four plans: Everyday Usage at $5 a month, High Usage at $15, Power Usage at $50 and Maximum Usage at $200. Meta’s model page lists Muse Wonder Full on the Power Usage and Maximum Usage plans only. The Muse personal agent app has a free tier with a usage limit. According to Meta’s help center, the paid plans are Power at $20 a month (500 million Muse tokens a week) and Maximum at $100 a month (3 billion a week). Our review of the Meta Muse AI agent covers what those plans do in practice. Maybe you only want to chat with the newest frontier models, without a new sign-up every time one launches. Fello AI puts GPT, Claude, Gemini, Grok and others in one Mac app, and new models usually arrive as soon as their API opens.

How to Use Muse Wonder Today

There are three ways to reach Muse Wonder on launch day, and only one of them includes Muse Wonder Full. Developers can call Muse Wonder through the Meta Model API, which works with OpenAI-SDK-compatible clients. OpenRouter also lists it, at the same $4 and $20 rates. Coders can use Muse Wonder inside Muse Code, and they get Muse Wonder Full only on the $50 Power Usage or $200 Maximum Usage plan.

The consumer route is less clear, because Meta describes it two ways. Meta’s launch post and its official account say Muse Wonder “will soon power experiences in Muse” and will roll out first to subscribers. Zuckerberg’s thread says “You can use Muse Wonder to power your Muse by subscribing today.” In practice, expect it on the $20 and $100 Muse plans first, and expect free users to wait. Zuckerberg wrote that Meta will “work to bring this level of intelligence to everyone over time,” and Meta’s post says it will use Wonder to build specialised models for everyone using Muse. Muse subscriptions are also still in limited testing and not available in every country.

What Meta’s Safety Card Discloses About Muse Wonder

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. The exact wording is “Without mitigations, we do not rule out the model reaching this threshold in the Chemical & Biological domain.” Meta says it applies safety training and system-level classifiers for that domain. With those in place, it rates residual risk at “moderate or lower” in every domain it tracks, which is the basis of its decision to ship.

The cyber results are worth reading closely. On ExploitGym, Muse Wonder Full captured 92 flags, but a verification pass found only 21 of them used the intended vulnerability. The other 71 came through unintended holes, so 21 is the figure comparable with the published ranking. Meta calls Muse Wonder “our most aligned model to date,” reports a 0% harmful-action rate on an agentic misalignment test, and says Muse Wonder is not intended for anyone under 18. A full safety report is promised “soon.” Until it lands, the card is the only safety document.

Is Muse Wonder Open Source?

No. Muse Wonder is a closed model, and neither Meta’s launch post nor its model card mentions releasing the weights. The only weights the card discusses are the ones Meta is protecting under its weight security program. Meta’s September promise has not moved either. On September 2, Zuckerberg wrote “Next up 🍉 and Muse Spark open weights releases coming soon.” The watermelon arrived. The Muse Spark weights have not, and Meta’s Hugging Face organisation still lists only the smaller Muse Glimmer 30B. The original Muse Spark launch was already closed, so Wonder continues the shift away from Meta’s open Llama era.

The Verdict on Meta Muse Wonder

Meta Muse Wonder puts Meta in the top tier of AI models, and the evidence comes from outside Meta, not just from its own chart. It is not the best model. Claude Opus 5.5 still leads the independent index. It wins more of the third-party rows on Meta’s scorecard, and it costs exactly the same per token as the Muse Wonder you can actually buy. The case for Muse Wonder is narrow and real: long autonomous coding projects and legal or document-heavy agent work, where Meta’s margins are large enough to survive its generous testing choices. If that is your workload, run it against Opus 5.5 on your own tasks this week. If it is not, there is no price reason to switch.

Frequently Asked Questions