Key Takeaways
- Muse Wonder Full is the most capable version of Meta’s new Muse Wonder model, launched on October 12, 2026, by Meta Superintelligence Labs. It handles long agentic tasks such as coding, research, planning and online errands across apps.
- Two new abilities set it apart: Active Memory, which keeps track of important details across very long sessions, and Active Focus, which keeps the model on track during multi-hour tasks.
- In Meta’s own comparison, Wonder Full posts the best score on legal, finance, automation and agentic coding tests, including DeepSWE (79.0), Harvey legal tasks (13.1) and Vals Finance Agent (59.1). Claude Opus 5.5 leads on terminal coding, computer use and health questions.
- Artificial Analysis scores Wonder Full 56 on its Intelligence Index, behind Claude Opus 5.5 at 58 but ahead of GPT-6 Astra, Gemini 4 Argon and Claude Fable 5.1 at 53.
- Wonder Full is only available through Muse Code Power ($50 a month) and Max ($200 a month) plans. The standard Muse Wonder model is available through Meta’s API at $4 per million input tokens and $20 per million output tokens.
- It accepts text, images, video, audio and PDFs, with a context window of 1,048,576 tokens.
Image source: Meta
Muse Wonder Full is Meta’s most capable AI model, built to carry out long, multi-step tasks on a user’s behalf. It can write and fix software over several hours, research a topic and act on the results, fill in forms and place orders across websites and apps, and plan events using information from photos, messages and documents.
Meta designed it around two new features, Active Memory and Active Focus, which help it remember what matters and stay on task during sessions that would make earlier models lose the thread.
Meta Superintelligence Labs released Muse Wonder on October 12, 2026, in two forms. The standard Muse Wonder is available to developers in preview. Wonder Full, a larger and more capable configuration, is reserved for the top tiers of Meta’s Muse Code subscription.
Meta calls the release “a major leap toward personal superintelligence,” and the benchmark results show a model that competes directly with the best from Anthropic, OpenAI and Google.
Muse Wonder vs Muse Wonder Full: what is the difference?
| Feature | Muse Wonder | Muse Wonder Full |
|---|---|---|
| Access | Meta Model API (preview), OpenRouter, all Muse Code plans | Muse Code Power ($50/month) and Max ($200/month) only |
| API model name | muse-wonder-1.0 | Not offered through the API |
| Inputs | Text, images, video, audio, PDF | |
| Output | Text | |
| Context window | 1,048,576 tokens | |
| Effort settings | Minimal to max | |
| Artificial Analysis Intelligence Index | 52 | 56 |
One point confuses many users: setting the standard Muse Wonder to maximum effort does not turn it into Wonder Full. Wonder Full is a separate, larger configuration.
Meta says its gains come partly from a bigger base model, and that improvements from post-training carry over better to unfamiliar tasks when the starting model is larger. In its research post, Meta shows Wonder Full outperforming the trend line set by Muse Spark 1.3 on out-of-distribution tests.
Active Memory and Active Focus Explained
Active Memory
Active Memory lets the model decide what to keep, update and recall as a long session unfolds, rather than treating everything in its context window as equally important.
Meta measures it with MemoryBench, where Muse Wonder scored 85.6 and Wonder Full 85.5, ahead of Claude Opus 5.5 at 81.5, GPT-6 Astra at 81.4 and Claude Fable 5.1 at 79.0. In practice, this means fewer cases where an agent forgets a constraint the user set an hour earlier.
Active Focus
Active Focus targets a different failure: drifting off course during very long tasks. Meta tests it with FocusBench and with SWE-Marathon, a set of extended software engineering jobs. On SWE-Marathon, Wonder Full scored 57.5 at extra-high effort, followed by Muse Wonder at 53.1, Claude Opus 5.5 at 46.3, Claude Fable 5.1 at 45.6 and GPT-6 Astra at 42.5. The gap is wider here than on most standard coding tests, which suggests Active Focus matters most when a task runs for hours.
Muse Wonder Full Benchmark Results
Meta compared Wonder Full with its own standard model and three rivals across ten measures. Wonder Full posts the top score on four, Claude Opus 5.5 on five, and GPT-6 Astra on one. Add Meta’s MemoryBench and SWE-Marathon results, covered above, and Meta’s models take two more.
| Benchmark | Wonder Full | Muse Wonder | Claude Fable 5.1 | GPT-6 Astra | Claude Opus 5.5 |
|---|---|---|---|---|---|
| Harvey LAB-AA (legal) | 13.1 | 10.0 | 6.4 | 8.6 | 4.2 |
| Vals Finance Agent V2 | 59.1 | 56.6 | 58.9 | 53.5 | 58.6 |
| AutomationBench-AA | 45.4 | 39.1 | 32.1 | 41.6 | 41.2 |
| DeepSWE | 79.0 | 78.8 | 67.4 | 74.1 | 74.2 |
| Terminal-Bench 4.0-AA | 57.1 | 52.0 | 52.0 | 59.1 | 59.6 |
| ProgramBench | 59.0 | 50.0 | 55.5 | 52.5 | 70.5 |
| OSWorld V2.1 (computer use) | 78.4 | 72.6 | 80.7 | 75.8 | 81.8 |
| HealthBench Pro | 62.3 | 60.5 | 62.1 | 65.3 | 65.6 |
| GDP.pdf | 33.2 | 30.8 | 27.6 | 34.2 | 30.6 |
| Artificial Analysis Index | 56 | 52 | 53 | 53 | 58 |
The pattern is clear. Wonder Full leads on legal, finance, business automation and agentic software engineering. Claude Opus 5.5 leads on terminal coding, ProgramBench, computer use, health questions and the overall Artificial Analysis score. GPT-6 Astra edges ahead on GDP.pdf, a test of document-based economic tasks.
What Muse Wonder Full Can Do in Practice
Meta’s launch demos lean toward everyday tasks rather than pure coding. In one, the model built a playable 3D kitchen game from a handful of reference images. In another, it handled an insurance claim from start to finish.
It planned a Saturday running route by working across a browser and several apps, ordered custom merchandise after reading Instagram photos, and produced a seating chart for a wedding. Meta also showed it offering wellness support in conversation.
These examples point to Meta’s strategy. While rivals focus heavily on coding agents, Meta is building toward personal agents that act across the apps people already use.
That approach puts Meta’s assets, from Instagram and WhatsApp to its smart glasses, at the center of the product. Our preview of what to expect from Meta AI in 2026 traced how this plan took shape, and the lab behind it is profiled in our look at Meta Superintelligence Labs versus the AI giants.
Where Muse Wonder Full Falls Short
Early users have reported weaker results than the benchmarks suggest on some tasks. In one widely shared Hacker News test, Wonder Full turned an image into HTML less accurately than Claude Opus 5.5, and several developers blamed problems with the Muse Code harness rather than the model itself.
Wonder Full also trails Anthropic’s models on terminal work and computer use, two areas where coding agents spend much of their time.
Developers deciding which tasks suit Claude Code or OpenAI Codex now have a third option to test, but not a clear replacement. Meta has also shown it will pull AI features that fail to land, as our report on why Meta axed the Muse Image feature on Instagram explains.
Access is the other limit. Wonder Full is not available through the API, so businesses cannot build products on it yet. Muse Code subscriptions require users to be 18 or older, vary by region, and come without refunds. At launch, Hacker News users reported the plans were limited to North America. Contributor mode, which lets Meta train on user traffic, is switched on by default for Muse Code and can be turned off in settings.
Safety Measures in Muse Wonder
Meta says Muse Wonder carries its strongest safeguards so far against misuse involving chemical, biological, radiological and nuclear threats, as well as offensive cyberattacks, while keeping defensive cybersecurity uses available.
These rules follow Meta’s Superintelligence Scaling Framework. The company describes Wonder as its most aligned model to date and has promised a full safety report soon.
How to Access Muse Wonder Full
Wonder Full is included in two Muse Code plans: Power at $50 a month and Max at $200 a month. The standard Muse Wonder model is available in every Muse Code plan, through Meta’s Model API in preview, and on OpenRouter. Meta says the Muse consumer app, which currently runs on Muse Spark 1.3, will get Muse Wonder soon, with paying subscribers first in line.
The launch landed well with investors and users. Meta’s stock rose about 3% after the announcement. The Muse app had already passed 1.8 million iOS downloads in the US and Canada during its first 12 days, according to Apptopia, ahead of ChatGPT’s 1.3 million over the same period after its launch.
For readers interested in building their own agent rather than renting one, our guide on how to develop a personal AI assistant covers the basics.
If you are interested in this topic, we suggest you check our articles:
- Meta Superintelligence Labs vs the AI Giants
- What Can We Expect From Meta AI in 2026?
- Why Meta Axed the Muse Image Instagram Feature
- Tasks Best Suited to Claude Code or OpenAI Codex
- How to Develop Your Own Personal AI Assistant
Sources: Meta – Muse Wonder, eesel AI – Muse Wonder, eesel AI – Muse Wonder pricing, YourStory, 36Kr
Written by Alius Noreika
