Meta Muse Wonder: A Frontier Agent, Not an Open-Weight Llama
Yoni Fraimorice
Meta is back in the flagship model race, but not with another open-weight Llama.
On October 12, Meta released Muse Wonder and Muse Wonder Full, two closed frontier models built for long-running agents, coding, computer use, and professional work.
The normal Muse Wonder is available in preview through the Meta Model API and Muse Code. The more capable Muse Wonder Full is available only through paid Muse Code subscriptions. Setting the API model to maximum reasoning does not turn it into Wonder Full.
That product split is the first thing developers need to understand.
What Muse Wonder actually is
According to Meta's model documentation, Muse Wonder accepts text, images, video, audio, and PDF files, then returns text. It has a 1,048,576-token context window and reasoning levels from minimal to max.
It is designed to:
- Plan multi-step work
- Delegate tasks to sub-agents
- Call tools in parallel
- Recover from errors
- Modify large codebases
- Use a computer interface
- Continue work in background responses
Meta calls two parts of this behavior Active Memory and Active Focus.
Active Memory means carrying useful information across conversations and retrieving it when it becomes relevant. That is different from a large context window, which only holds information inside one request or session. Meta does not publish the full memory architecture, so developers should not assume where or how personal data is stored.
Active Focus means persisting through blockers, changing approaches, and recovering from failures. Meta says the model should ask for clarification and check before important actions.
Access and price
Muse Wonder's API model ID is muse-wonder-1.0. In Standard mode, Meta says prompts, files, context, and outputs are not used to train its models.
- $4 per million input tokens
- $20 per million output tokens
- $0.20 per million cached input tokens
There is no long-context surcharge.
Those token prices exactly match Claude Opus 5.5. GPT-6 Astra costs $10 per million input tokens and $50 per million output tokens.
Muse Wonder Full has no pay-as-you-go API. It is a different, higher-capability variant available only inside the Power Usage and Maximum Usage Muse Code plans.
Muse Code subscriptions default to Contributor mode, which gives more usage in exchange for permission to use submitted content to train future Meta models. Users can turn it off. Teams working with private source code should check this setting before the first session.
The benchmark comparison is mixed
Meta's Muse Wonder Model Card compares both variants with Claude Fable 5.1, GPT-6 Astra, and Claude Opus 5.5. All are shown at maximum reasoning effort.
Selected results:
| Evaluation | Wonder Full | Wonder | GPT-6 Astra | Opus 5.5 |
|---|---|---|---|---|
| Personal Superintelligence Index | 59.1 | 54.3 | 54.5 | 55.2 |
| AutomationBench-AA | 45.4 | 39.1 | 41.6 | 41.2 |
| DeepSWE 1.1 | 79.0 | 78.8 | 74.1 | 74.2 |
| SWE-Marathon | 57.5 | 53.1 | 42.5 | 46.3 |
| OSWorld V2.1, partial score | 78.4 | 72.6 | 75.8 | 81.8 |
| ProgramBench | 59.0 | 50.0 | 52.5 | 70.5 |
| Artificial Analysis Index | 56 | 52 | 53 | 58 |
The pattern is clear:
- Wonder Full leads Meta's long-horizon agent and coding rows.
- Opus 5.5 remains stronger on the broad external index, computer use, and clean-room program reconstruction.
- The API-accessible Wonder is close to Full on DeepSWE, but the gap is larger on professional work and automation.
Meta's methodology also needs a warning label. Some competitor results came from provider reports or external leaderboards. Other comparisons used Meta's internal harness. Meta says those competitor runs are best-effort and may not be optimally tuned.
The Personal Superintelligence Index is also Meta's own composite, built from seven agent and knowledge-work evaluations. It is useful, but it is not an independent universal ranking.
How it compares with Gemini 4 Argon
Google's Gemini 4 Argon announcement reports 77.9% on DeepSWE 1.1 and 51.3% on AutomationBench.
Muse Wonder Full reports 79.0% on DeepSWE and 45.4% on AutomationBench-AA. This suggests similar frontier coding strength, while Argon may be stronger on end-to-end business automation.
But these are not a clean head-to-head test. The sources, harnesses, and even the AutomationBench variants differ. Treat the numbers as direction, not a final league table.
How it compares with Opus 5.5
Muse Wonder and Claude Opus 5.5 share a 1M-token context window and the same base token price.
Opus 5.5 accepts text and images. Muse Wonder adds video, audio, and PDF input through one model endpoint. Opus is available through Anthropic, AWS, Google Cloud, Microsoft Foundry, and Claude Platform on AWS; Muse Wonder currently centers on Meta's own API and Muse Code.
For coding agents, the choice is not obvious. Wonder Full leads Meta's DeepSWE and SWE-Marathon table. Opus 5.5 leads ProgramBench and OSWorld, has a mature Claude Code workflow, and Anthropic documents action screening and sandboxing around the agent.
How it compares with GPT-6 Astra
GPT-6 Astra has a slightly larger 1.05M-token context window, the same 128K maximum output class, and a broad hosted tool set including web search, code interpreter, shell, computer use, MCP, and image generation.
Astra is more expensive, but OpenAI positions it for the most demanding reasoning, research, coding, and computer-use work. In Meta's table, Wonder Full wins the selected long-horizon coding and automation rows, while Astra is stronger on GDP.pdf and Terminal-Bench.
Cost may decide many workloads: Muse Wonder's $4/$20 token pricing is 60% below Astra's $10/$50 rates before tool fees and caching.
Safety: strong claims, incomplete certainty
Meta's model card evaluates the Full variant as the most capable member of the family. It says the unmitigated model could reach Meta's high-risk threshold in chemical and biological capability, so Meta applies safety training and system classifiers. With mitigations, Meta rates residual risk as moderate or lower across CBRN, cybersecurity, and loss of control.
The card also reports strong refusal performance and a 0% harmful-action rate on its Agentic Misalignment evaluation.
These are Meta's evaluations, not proof that every deployment is safe. The card warns that multi-step reasoning can still fail in unfamiliar scenarios and recommends human confirmation for irreversible actions.
Active Focus makes that advice important. Persistence should have a budget, a stop condition, and an approval boundary.
Who should use it?
Choose Muse Wonder when you want a multimodal, long-context agent through an API at Opus-level token pricing.
Choose Wonder Full when Muse Code fits your workflow and maximum long-horizon performance matters more than API portability.
Keep Opus 5.5 on the shortlist for mature coding-agent workflows, strong computer use, and broad cloud availability.
Use GPT-6 Astra when you need OpenAI's tool ecosystem or its highest reasoning tier and can justify the higher price.
Consider Gemini 4 Argon for Google-centered enterprise and cybersecurity workflows.
Then run your own evaluation. At this level, the benchmark winner often changes with the prompt, tool harness, repository, and stopping rules.
Sources
- Meta: Introducing Muse Wonder and Muse Wonder Full
- Meta Research: Muse Wonder technical overview
- Meta: Muse Wonder Model Card
- Meta Model API: Models
- Meta Model API: Pricing and rate limits
- Anthropic: Claude Opus 5.5
- OpenAI API: GPT-6 Astra
- Google: Gemini 4 Argon
Hero photo: Meta Platforms Headquarters Menlo Park California, photographed by Wikimedia Commons user LPS.1 on March 27, 2022. CC0 1.0 public-domain dedication. Resized from 4512 x 3008 to 1920 x 1280 pixels and JPEG-compressed; no other edits. This illustrative photograph shows the entrance to Meta's headquarters. It does not show Muse Wonder, its training infrastructure, or a benchmark run. No endorsement is implied.