Meta's Muse Wonder: A First Frontier Model, With Memory as Its Selling Point

Until this week, Meta’s AI story since the 2025 reorganisation under Alexandr Wang had two parts: a fast, cheap model family (Muse Spark), and a consumer agent (Muse) built on it. What it lacked was a model that could sit at the same table as Anthropic, OpenAI and Google. On Monday, 12 October 2026, Meta’s Superintelligence Lab released two of them: Muse Wonder and Muse Wonder Full. TechCrunch called it Meta’s first frontier model. Wang put it more bluntly, calling Muse Wonder “the best model in the world for personal agents and professional agents.”

Independent numbers are already available, and the picture they give is more mixed than the launch claims, though still a real step up.

What was released

Two variants. Muse Wonder is available now through the Meta Model API (dev.meta.ai) and the Muse Code coding product. Muse Wonder Full, the stronger variant, is only available to Muse Code subscribers and is not served through the API.
Price. Muse Wonder costs $4 per million input tokens and $20 per million output tokens, according to both TechCrunch and The Decoder. That is the same bracket as GPT-6 Sol and Claude Opus 5.5, and a big jump from Muse Spark 1.3’s $1.25 / $4.25. Wonder Full has no published token price.
Specs. The model card lists a 1M-token context window, input of text, images, video and audio, and text-only output. Neither the blog post nor the model card mentions open weights. This is an API-only release, a clear break from the Llama era.
Training. Meta says both models were trained at its gigawatt-scale Prometheus data center in Ohio, using pretraining plus reinforcement learning on top of the Muse Spark scaling work.
Rollout. Muse Wonder will power the Muse agent app, paying subscribers first. Meta says it will later distil specialised models from it for everyone else.

What is actually new: memory and persistence

Meta is not pitching Wonder mainly on raw reasoning. It pitches two behaviours aimed at agents:

Active Memory. The model is meant to carry important context about a user across conversations and pull in the right earlier detail when a new task needs it. Meta measures this with MemoryBench, an internal benchmark that checks recall of key facts buried in 64K–100K-token multi-turn conversations. These conversations include distractor threads and injected turns that contradict or update the target facts. According to TechCrunch, Meta reports that Wonder beats Claude Fable 5.1, Opus 5.5 and GPT-6 Astra on it.
Active Focus. The model keeps working on hard goals by trying alternative approaches, working around blockers, asking for clarification and checking in before important actions. The internal FocusBench covers personal tasks in housing, travel, IT and finance where the environment changes mid-task.

Both behaviours matter for an agent that will have access to your calendar, messages and documents, which is where Meta plans to deploy it. But both headline benchmarks are internal to Meta. There is no independent replication yet, and it is easier to win on a test you designed yourself.

The independent numbers

Artificial Analysis, which Meta itself cites in its model card, gives a less flattering but still impressive picture:

Intelligence Index v4.3: Wonder Full 56, Wonder 52, Muse Spark 1.3 48. On the Artificial Analysis model page, Wonder Full (Max) ranks 4th of roughly 700 models listed.
Coding Agents Index: Wonder Full 68, tying Claude Code running Sonnet 5.5. Wonder scores 63, against 54 for its predecessor (The Decoder).
Terminal-Bench 4.0: up from 33% to 57% over the predecessor, with about 30% fewer output tokens for Wonder Full.
ProofBench (private math proofs): Wonder Full solved 94/100, up from 58. The leaders score 99–100.
Factual knowledge is the weak spot. On AA-Omniscience, Wonder Full scores 30 and Wonder 25, against 46 for Opus 5.5, 43 for GPT-6 Astra and 42 for Gemini 4 Argon. Its hallucination rate fell from 33% to 23%, which is good but still above Gemini 4 Argon’s 15%.
Cost per task on the index is $4.52 for Wonder, nearly three times Muse Spark 1.3’s $1.60. That is cheaper than Opus 5.5 ($5.98) and Fable 5.1 ($7.63), but more than GPT-6 Astra ($3.26) and Gemini 4 Argon ($1.99). Input tokens make up about 76% of the bill, and Wonder uses 28% more of them per task than its predecessor.
On the Vals AI Index, Wonder Full (Max) ranks fifth, ahead of every OpenAI model but behind current Anthropic models.

Read the model card’s fine print

Meta’s own benchmark table deserves some care. The methodology section says that where no third-party result exists, Meta uses “the highest of our internally validated run, self-reported score, or official leaderboard score”. It also admits that its runs of competitor models “may not be optimally tuned for proprietary models”. The SWE-Marathon row compares Wonder Full against GPT-6 Astra inside Codex and Fable 5.1 inside Claude Code. The card itself says this is not a controlled common-harness comparison. None of this is unusual for a vendor model card, but it means the self-reported table is a ceiling, not a neutral measurement.

The safety section has one detail developers working on security should notice. On ExploitGym, Wonder Full captured 92 flags, but a verification pass found that only 21 used the intended vulnerability. The other 71 came through unintended vulnerabilities in the test environments. Meta reports 21 as the comparable figure. Another reading is that the model is very good at finding the bug nobody planted. That is useful for defenders, and it is the same kind of behaviour that has recently caused trouble for other labs’ agents in live environments. Under its Superintelligence Scaling Framework, Meta says it could not rule out the unmitigated model reaching its high-risk threshold for chemical and biological misuse. It therefore shipped with safety post-training and system-level classifiers, and rates residual risk “moderate or lower”. A full safety report has been promised “soon”.

Why it matters

There are three reasons.

First, the frontier now has four serious labs, not three. A credible Meta flagship adds price pressure on the $4–5/M-input tier and gives enterprise buyers another option, especially those already tied to Meta’s ad and messaging platforms.

Second, Meta is betting that the next competition is about persistent personal agents, not chat quality. Memory across sessions and the willingness to keep going after a blocker are the two things today’s coding and browsing agents most often lack. If MemoryBench-style gains hold up outside Meta’s lab, other vendors will have to answer.

Third, for developers this is a closed, API-only Meta model. Anyone who built on Llama’s open weights should notice that Meta’s best work now sits behind dev.meta.ai, and that the strongest variant is not on the API at all.

Open questions

Independent memory tests. Until outside groups test long-horizon recall, including what happens with stale or conflicting memories, “Active Memory” is a marketing claim with an internal benchmark behind it.
Knowledge gap. An index score near the top combined with an AA-Omniscience score 16 points behind Opus suggests strong reasoning on a weaker base of knowledge. The Decoder speculates the base model is smaller than its rivals’. Meta has not said.
Benchmark tuning. Earlier Muse models were criticised for scores that did not carry over to real use. Meta says Wonder Full beats the trend line on out-of-distribution tasks, but that also comes from its own testing.
Wonder Full pricing and API access are still unknown, which makes the more impressive half of the launch hard to evaluate in production.

The fair summary: Meta has caught up enough to matter, has not taken the lead, and is competing on memory and persistence rather than raw knowledge. Whether that is the right bet depends on numbers nobody outside Meta has produced yet.