Meta Muse Wonder Explained: Active Memory, Active Focus, and the Hype
Meta's Muse Wonder and Muse Wonder Full explained: what active memory and active focus do, how benchmarks stack up, and where it falls short.
What is Meta Muse Wonder?
Muse Wonder is Meta’s first model that the company itself calls “frontier class,” released on October 12 by Meta Super Intelligence Labs alongside a larger sibling called Muse Wonder Full. Muse Wonder is available through Meta’s model API and inside Muse Code, Meta’s terminal based coding agent. Muse Wonder Full, the bigger and reportedly stronger of the two, is not available via API at all. It only ships through a Muse Code subscription on the $50 or $200 monthly plans. Both were trained at Meta’s Prometheus data center in Ohio, and Muse Wonder is slated to power the standalone Muse app, rolling out to subscribers first.
TL;DR
- Meta shipped two models at once: Muse Wonder (API accessible) and Muse Wonder Full (locked behind a Muse Code subscription), both described as frontier class.
- The headline features are active memory, which retains context about a user across days, weeks, and months, and active focus, which trains the model to persist through blockers instead of stopping and handing a task back.
- On Meta’s own benchmarks and on independent trackers like Artificial Analysis, Muse Wonder Full lands close to Claude Opus 5.5, generally a few points behind it and ahead of GPT6 Astra on several measures.
- Meta is explicitly positioning the model around personal and professional agents rather than chatbot leaderboards, with demos covering gift design, rental hunting, running plans, and wedding seating charts.
- Independent testers found a gap between the benchmark story and early demos, with 3D rendering and game-building outputs falling short of what Opus 5.5 and GPT 6.1 Soul have produced.
- Meta went from a rough Llama 4 launch and a benchmark controversy in April 2025 to a model competing with the top labs in roughly 14 months, after a major reorganization around Alexander Wang’s Meta Superintelligence Labs.
- Usage has been rocky at launch, with reports of quota errors (429s) even on the $200 a month plan, which Alexander Wang acknowledged as a known issue Meta was actively fixing.
One coffee. One working app.
You bring the idea. Remy manages the project.
What do active memory and active focus actually do?
Active memory is Meta’s term for persistent context retention. Instead of starting fresh in every conversation, the model is meant to hold onto details that matter (preferences, past requests, ongoing projects) and surface them when relevant, without the user re-explaining themselves. Meta describes it as working “over days, weeks, months, and beyond,” covering everything from small habits to the substance of earlier conversations.
Active focus addresses a different failure mode: agents that do fine on the first few steps of a task but stall the moment something unexpected happens. Meta says Muse Wonder was trained specifically not to give up when it hits a blocker, instead trying alternative approaches to still complete the task. Combined, these two features are the foundation of Meta’s pitch that this isn’t a model built to win trivia-style benchmarks, but one built to carry out multi-step, long-running work on someone’s behalf.
How does Muse Wonder perform on benchmarks?
Meta’s own comparisons pit Muse Wonder and Muse Wonder Full against Claude Opus 5.5, Claude Fable 5.1, and GPT6 Astra. On Meta’s internal Personal Super Intelligence Index, a composite covering personal and knowledge work, Muse Wonder Full scored 59.1%, ahead of Opus 5.5 (55.2%), GPT6 Astra (54.5%), and Fable 5.1 (49.9%). On Automation Bench, which measures end-to-end business workflows, Wonder Full scored 45.4%, topping the chart. On Harvey’s legal agent benchmark it scored 13.1% against 4.2% for Opus 5.5, though a separate source noted Grok 4.7 scored higher (19.6%) without appearing on Meta’s chart.
On long-horizon coding tasks, Muse Wonder Full scored 79% on Deep SWE 1.1 and Muse Wonder scored 78.8%, both ahead of Opus 5.5 (74.2%) and GPT6 Astra (74.1%). On SWE Marathon, described as ultra-long-horizon engineering, Wonder Full scored 57.5% versus 46.3% for Opus 5.5.
Independent verification came from Artificial Analysis, whose intelligence index (a composite of ten evaluations including GPT Val, Terminal Bench, SciCode, Humanity’s Last Exam, and AA Omniscience) placed Muse Wonder Full at 56, just behind Opus 5.5 at 58 and tied with Claude Sonnet 5.5. Regular Muse Wonder scored 52, near GPT6 Astra and GPT 6.1 Soul. On the Vals index focused on real professional tasks, Muse Wonder Full ranked fifth at 64.57% accuracy, ahead of every OpenAI model tested, which led Vals to publish a post titled “Meta now outranks OpenAI on the Vals index.”
On OS World, a benchmark for controlling a computer or browser, Opus 5.5 still led (81%) ahead of Fable 5.1 (80%), with Muse Wonder Full at 78.4% and Muse Wonder at 72.6%.
Do the real-world demos match the benchmark claims?
This is where the launch got messier. Independent testing surfaced a noticeable gap between Meta’s benchmark charts and what the model actually produced in creative and rendering tasks. One tester’s “hyperrealistic” 3D boat render on Three.js showed visible artifacts: odd white residue around the boat, poor shadow rendering, and a sun that looked out of place despite working time-of-day and wind controls. A Fall Guys-style game build showed flickering and clipping issues. A hyperrealistic watermelon render, built as a nod to the model’s reported codename, turned out blurry rather than photorealistic. By comparison, testers noted Opus 5.5 and GPT 6.1 Soul produced cleaner results on similar prompts.
Not every test underperformed. A Rubik’s Cube solving demo worked convincingly, with the model scrambling and then genuinely solving the cube (not just reversing its own moves) in about four minutes. One tester also received an unusual, seemingly unprompted output during a render task warning against investigating its “origin, ancestry, or self-inflicted characterization,” which had no apparent connection to the chain of thought and drew attention online simply for being strange.
The broader criticism forming around the launch is the idea of “benchmaxing”: training a model heavily toward specific test formats so benchmark scores look stronger than the model’s general capability. Meta has history here. When Llama 4 launched in April 2025, the version of Llama 4 Maverick submitted to the LMArena leaderboard turned out to be a special variant tuned for human preference, not the model people could actually download. LMArena publicly said Meta’s interpretation of its policy didn’t match what it expected from model providers. That episode is part of why some of the Muse Wonder benchmark results are being read skeptically rather than taken at face value.
Is Muse Wonder worth using right now?
Pricing puts Muse Wonder in a competitive but not cheap bracket. It runs $4 per million input tokens and $20 per million output tokens, with cached input at 20 cents per million. That’s roughly double GPT 6.1 Soul’s short-context pricing ($2 input, $10 output, 10 cents cached) but close to Opus 5.5. On cost-per-completed-task metrics, Muse Wonder Max came in around $4.52 per task, similar to the open-source Qwen 38, while GPT 6.1 Soul Max was reported as more expensive per task despite being favored by at least one reviewer for overall value given its quota and intelligence tradeoffs.
Access has also been rough at launch. Testers reported hitting usage quotas (429 errors) within minutes even on the $200 a month Muse Code plan, before discovering that lower tiers didn’t include the full model at all. Alexander Wang responded directly to at least one public complaint, confirming the issue was known and being fixed.
For developers, the practical read is that Muse Wonder (not Full) is the only version reachable via API today, and it’s priced and benchmarked competitively with GPT6 Astra rather than clearly ahead of it. Muse Wonder Full, the stronger model, currently requires a paid Muse Code subscription and has shown both genuinely strong coding benchmark results and inconsistent creative/rendering output in early hands-on testing.
FAQ
What is the difference between Muse Wonder and Muse Wonder Full?
Muse Wonder is accessible through Meta’s model API and Muse Code. Muse Wonder Full is the larger, more capable version, but it’s only available through a Muse Code subscription on the $50 or $200 monthly plans, not via API.
What are active memory and active focus in Muse Wonder?
Seven tools to build an app. Or just Remy.
Editor, preview, AI agents, deploy — all in one tab. Nothing to install.
Active memory lets the model retain and recall relevant context about a user over long periods (days to months). Active focus is training aimed at persistence, so the model tries alternative approaches when it hits a blocker instead of stopping.
How does Muse Wonder compare to Claude Opus 5.5 and GPT6 Astra?
On Artificial Analysis’s intelligence index, Muse Wonder Full scored 56 versus Opus 5.5’s 58, putting it just behind the top model and ahead of GPT6 Astra on several individual benchmarks, including Deep SWE 1.1 and the Vals professional-task index.
Is Muse Wonder benchmaxed?
Some reviewers and commentators have raised the possibility, partly because of Meta’s history with Llama 4’s LMArena submission controversy and partly because early hands-on demos (3D rendering, game building) underperformed relative to the benchmark scores.
Can I use Muse Wonder Full through an API?
No. As of launch, Muse Wonder Full is only accessible through a Muse Code subscription on paid tiers, not through Meta’s public API.
