The Announcement of Meta's First Frontier Models
Meta Superintelligence Labs presented Muse Wonder and Muse Wonder Full on October 12, 2026, two models the company describes as its most capable systems to date. This launch marks a strategic inflection point, as they are the first Meta models that the company explicitly equates in performance to those of OpenAI and Anthropic, placing them in what the industry calls the 'frontier' of artificial intelligence.
The initiative follows a deep restructuring initiated fourteen months ago, which included hiring Alexandr Wang, former head of Scale AI, to lead the AI division. The stated goal is to compete directly in the development of high-performance models for personal and professional agentic tasks, moving away from the previous strategy centered on open-source models like the Llama family.
Differences Between Muse Wonder and Muse Wonder Full
The main distinction between the two models lies in their availability and access channel. Muse Wonder is designed for developers and is immediately available through the Meta Model API and the Muse Code tool. On the other hand, Muse Wonder Full is a more powerful version offered exclusively to Muse Code subscribers, with no direct API access at the moment.
Meta has not exhaustively detailed the internal technical differences between the two, but presents Wonder as the foundation for future specialized models that will be integrated into the Muse application. The Full variant is positioned as the highest-capacity option for users requiring a professional-level AI-assisted work environment, especially in programming.
Availability and Access Comparison
Summary of the differences confirmed by Meta in the official announcement.
| Feature | Muse Wonder | Muse Wonder Full |
|---|---|---|
| Primary Access | Meta Model API and Muse Code | Muse Code subscribers only |
| API Pricing | $4/M input tokens, $20/M output tokens | Not available via API |
| Integration in Muse | Coming soon for subscribers | Included in Muse Code subscription |
Technical Capabilities: Active Memory and Active Focus
Meta highlights two key features that define the behavior of these models. The first is 'Active Memory', a capability that allows the system to remember and maintain context across multiple conversations, retrieving relevant details when needed for the current task. This aims to prevent users from having to constantly reconstruct information when resuming a project.
The second feature is 'Active Focus', designed for the model to work autonomously on complex, long-duration tasks. This capability allows the system to maintain focus on the goal, try alternative approaches when encountering errors, recover from failures, and request clarifications before executing important actions. Meta claims these features allow the generation of 'artifacts', such as functional web applications, and the management of workflows in law, finance, and health.
Performance and Evaluations: Claimed vs. Verified
Meta attributes competitive performance to its models in the Artificial Analysis index, placing Muse Wonder Full in second position, behind Anthropic's Claude Opus 5.5. The company claims to outperform OpenAI's GPT-6 Astra and Claude Fable 5.1 in memory tests (MemoryBench) and specialized software engineering tasks (SWE-Marathon, DeepSWE 1.1).
However, it is crucial to distinguish between Meta's claims and independent verifications. According to data cited in the coverage, Claude Opus 5.5 remains ahead in the general Artificial Analysis index and in programming tests like Terminal-Bench. GPT-6 Astra leads in professional work and health (HealthBench Pro). Therefore, although Meta has closed the gap in several metrics, it has not generally surpassed all its rivals in all tests.
Relative Position According to Cited Evaluations
Comparison of positions in specific indices according to data attributed to Meta and Artificial Analysis.
| Index / Test | Leader | Muse Wonder / Full |
|---|---|---|
| Artificial Analysis Index | Claude Opus 5.5 (58 pts) | Muse Wonder Full (56 pts) |
| MemoryBench | Muse Wonder | Outperforms Fable 5.1, Opus 5.5, and Astra |
| Terminal-Bench 4.0 | GPT-6 Astra | Behind Astra |
| HealthBench Pro | GPT-6 Astra | Behind Astra |
Availability, Pricing, and Security
The price of Muse Wonder via the API has been set at $4 per million input tokens and $20 per million output tokens. This rate positions it in the same cost range as GPT-5.6 Sol and Claude Opus 5.5, confirming its status as a high-performance, paid model. For end users, access to Wonder's capabilities in the Muse application will be linked to subscription plans of $20 and $100 per month.
Regarding security, Meta claims that Muse Wonder is its 'most aligned model to date', trained to reject dangerous requests and protect user privacy. The company has announced that a full report with additional safety evaluations will be available soon, although it has not yet detailed the specific safeguards implemented against misalignment or misuse.
How to Access Muse Wonder
Steps to integrate or use the new models according to user profile.
- 01Developers: Request access to the Meta Model API to integrate Muse Wonder into their own applications.
- 02Programmers: Subscribe to Muse Code to use Muse Wonder or Muse Wonder Full in the AI-assisted development environment.
- 03End Users: Await integration into the Muse app, initially available for paid subscribers ($20 or $100/month).
Open questions
- The full security and alignment report promised by Meta is yet to be published.
- No complete independent evaluations have been conducted to confirm Meta's performance claims outside the cited indices.
- Specific technical differences between Muse Wonder and Muse Wonder Full have not been detailed beyond access and general capability.
Keep exploring
Sources consulted
Corrections and transparency
If you spot incorrect or outdated information, send us a correction with the page and source we should review.
Submit a correction