Meta finally has a model that belongs at the frontier table. Whether it actually belongs at the head of that table is a different question.
On October 12, 2026 (UTC), Meta Superintelligence Labs dropped the Muse Wonder family — two models, Muse Wonder and Muse Wonder Full, codenamed Watermelon inside the company. Fourteen months after Mark Zuckerberg declared a ground-up overhaul of Meta's AI work and hired Scale AI founder Alexandr Wang as chief AI officer, the company is shipping what it calls "our most capable models to date."
What Meta actually built
Muse Wonder is not a general-purpose chat model. It is built for agents — the kind that book your dinner, negotiate a medical bill, or build a working app over several hours. Two features define the family:
- Active Memory carries context about you across conversations and pulls up the right detail when it becomes relevant. Meta measures this with an internal test called MemoryBench, where both Wonder models score around 85.5.
- Active Focus keeps the agent working when a task hits a wall — trying another approach, recovering from errors, checking in before doing something important. Wang told CNBC the goal was an agent that doesn't stall when a dinner booking hits a snag.
Both models take a 1 million token context window and accept text, images, video, and audio as input. They answer in text only. Training happened at Meta's gigawatt-scale Prometheus data center in central Ohio.
The availability split matters. Muse Wonder is on the public Meta Model API today and in Muse Code. Muse Wonder Full — the larger, stronger version — is locked inside Muse Code on the $50 and $200 monthly plans. You cannot buy it by the token.
The benchmark story, with caveats
Meta published a 14-row scorecard comparing both Wonder models against Claude Fable 5.1, GPT-6 Astra, and Claude Opus 5.5, all at max reasoning effort. Here is the full table:
| Benchmark | Wonder Full | Wonder | Fable 5.1 | GPT-6 Astra | Opus 5.5 |
|---|---|---|---|---|---|
| Personal Superintel. Index (internal) | 59.1 | 54.3 | 49.9 | 54.5 | 55.2 |
| Harvey LAB-AA (legal agents) | 13.1 | 10.0 | 6.4 | 8.6 | 4.2 |
| Vals Finance Agent V2 | 59.1 | 56.6 | 58.9 | 53.5 | 58.6 |
| HealthBench Pro | 62.3 | 60.5 | 62.1 | 65.3 | 65.6 |
| AutomationBench-AA | 45.4 | 39.1 | 32.1 | 41.6 | 41.2 |
| JobBench | 75.3 | 70.5 | 73.1 | 66.1 | 75.4 |
| OSWorld V2.1 | 78.4 | 72.6 | 80.7 | 75.8 | 81.8 |
| MemoryBench (internal) | 85.5 | 85.6 | 79.0 | 81.4 | 81.5 |
| DeepSWE 1.1 (long coding) | 79.0 | 78.8 | 67.4 | 74.1 | 74.2 |
| SWE-Marathon (very long projects) | 57.5 (xhigh) | 53.1 | 45.6 | 42.5 | 46.3 |
| ProgramBench (from scratch) | 59.0 | 50.0 | 55.5 | 52.5 | 70.5 |
| Artificial Analysis Index | 56 | 52 | 53 | 53 | 58 |
Source: Meta's published scorecard, October 12, 2026. Best score per row in bold.
Wonder Full wins seven of the twelve rows with data, but two of those wins are on Meta's own internal benchmarks (Personal Superintelligence Index and MemoryBench). Strip those out and Opus 5.5 wins six of the remaining ten to Wonder Full's four. The SWE-Marathon headline uses Wonder Full at xhigh effort while the rest of the table runs at max, and Meta itself flags that row as "not a controlled common-harness comparison."
Where Wonder Full genuinely leads is long-horizon coding and legal work. On SWE-Marathon it scores 57.5 against Opus 5.5's 46.3 — a 24% gap. On Harvey's legal agent benchmark it more than triples Opus 5.5, 13.1 to 4.2, though every model completes only a small share of tasks on that benchmark.
Independent testing tells a more sober story. Artificial Analysis Intelligence Index v4.3.2 scores the models like this:
| Model | AA Index v4.3.2 | Public API? |
|---|---|---|
| Claude Opus 5.5 (max) | 58 | Yes |
| Claude Sonnet 5.5 (max) | 56 | Yes |
| Muse Wonder Full (max) | 56 | No, Muse Code only |
| Claude Fable 5.1 (max) | 53 | Yes |
| GPT-6 Astra (max) | 53 | Yes |
| Muse Wonder (max) | 52 | Yes |
Source: Artificial Analysis, v4.3.2, October 2026.
Wonder Full ties Sonnet 5.5 at 56, two points behind Opus 5.5. The public API version, Muse Wonder, scores 52 — one point below both Fable 5.1 and GPT-6 Astra.
The pricing problem
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.
Wang told CNBC the model would be priced "meaningfully below" OpenAI and Anthropic. That is true against Fable 5.1 ($10/$50) and GPT-6 Astra ($10/$50). It is not true against Opus 5.5, which costs exactly the same and scores six points higher on the independent index.
The harder comparison is Claude Sonnet 5.5: at $2/$10, half Muse Wonder's price, it scores 56 on the same index — four points above the Muse Wonder you can actually buy.
| Model (API) | Input / M | Cached | Output / M | AA Index |
|---|---|---|---|---|
| Muse Wonder | $4.00 | $0.20 | $20.00 | 52 |
| Claude Opus 5.5 | $4.00 | $0.20 | $20.00 | 58 |
| Claude Sonnet 5.5 | $2.00 | $0.10 | $10.00 | 56 |
| Claude Fable 5.1 | $10.00 | $0.25 | $50.00 | 53 |
| GPT-6 Astra | $10.00 | $1.00 | $50.00 | 53 |
Source: Model pricing pages and Artificial Analysis, October 2026.
Why it matters
This is the first time Meta has shipped a model that independent testers put in the same tier as the frontier. For fourteen months, Meta's AI output was Muse Spark — a capable but clearly second-tier model that scored well below GPT-5.x and Claude Opus on every major benchmark. Wonder Full changes that. It is not the best model, but it is in the conversation, and on long coding projects and legal agent work it has a real lead.
The strategic significance is that Meta now has a model strong enough to power its consumer agent ambitions. Muse, the personal AI agent Meta launched on September 8, 2026 (UTC), climbed to the top of the US App Store free chart within weeks. It was running on Muse Spark. Upgrading it to Wonder — which Meta says will happen "soon," first for subscribers — closes the capability gap with ChatGPT and Claude in the one area where agents actually matter: getting things done over long horizons.
The benchmark margins deserve a skeptical read. Meta's own chart is the one showing Wonder Full ahead, and it includes two internal benchmarks plus a SWE-Marathon row run at a higher effort setting than the competitors. The independent index, which Meta did not control, puts Opus 5.5 clearly on top. This is not a GPT-6 moment where a new model sweeps every benchmark. It is a credible entry from a company that was, until now, watching the frontier from the sidelines.
The closed-weight decision is the biggest tell. Meta built its reputation in AI on open sourcing Llama. The Muse family has been closed from the start, and the promised Muse Spark open weights from September have not shipped. Wonder continues that pattern. Zuckerberg's company has decided that frontier capability and open weights are incompatible — at least for now. That is a philosophical shift with real competitive consequences: if Meta stops feeding the open-source ecosystem, the gap between open and closed models widens, and the companies relying on Llama as a free base layer have to reconsider their stack.
What to watch
- Muse app upgrade timing. Meta says Wonder will power Muse "soon," first for subscribers. The speed of that rollout tells you whether Wonder is production-ready or still needs hardening.
- Independent coding benchmarks. SWE-Marathon and DeepSWE are where Wonder Full's lead is largest. Watch whether third-party harnesses like LiveBench and Terminal-Bench confirm that lead or compress it.
- Pricing response from Anthropic. If Wonder takes coding share at Opus 5.5's price point, Anthropic may cut Opus pricing or push Sonnet harder as the value option.
- Open weights (or the lack thereof). Meta has not committed to releasing Wonder weights. If it never does, the Llama open-source era is effectively over at the frontier.
Meta is back in the frontier race. The question was never whether it could spend the money — it was whether it could build the model. With Muse Wonder, the answer is yes, with an asterisk. The asterisk is that the model you can actually buy is not the model that wins the benchmarks, and the model that wins the benchmarks costs the same as a better one from Anthropic.
No comments yet