Meta Superintelligence Labs has released Muse Wonder and Muse Wonder Full, the company's first models built to compete at the frontier with OpenAI, Anthropic and Google. Announced on October 12, 2026, the pair score close to the industry's best on independent rankings and lead on several agentic coding benchmarks in Meta's own tests. But in a sharp break from the open Llama era, neither model ships with downloadable weights, and the stronger of the two is not even available through Meta's API.
A Larger Class of Muse Models
The Wonder models sit above Muse Spark, the lightweight, low-latency model Meta introduced in April 2026. Mark Zuckerberg described Wonder as a much larger and more advanced class of model, and Meta's launch post framed the pair as the next step on its "scaling ladder" toward what it calls personal superintelligence.
According to VentureBeat, the key technical specifications are:
- Multimodal input: text, images, audio and video, with text output
- Context window: roughly 1 million tokens
- Training: large-scale pretraining plus reinforcement learning, run at Meta's Prometheus data center in central Ohio, which came online earlier this year
This is the first frontier release since Alexandr Wang restructured Meta's AI organisation in 2025. Since then, the lab had shipped only Muse Spark, image and video models, and the Muse agent, so Wonder is widely being read as the first real test of what the reorganised team can deliver.
How the Benchmarks Stack Up
On the independent Artificial Analysis Intelligence Index, Wonder Full scores 56, tying Claude Sonnet 5.5 and sitting two points behind Claude Opus 5.5 at 58. Standard Wonder lands at 52, just below GPT-6 Astra, Gemini 4 Argon and Claude Fable 5.1, which each score 53.
Meta's own benchmark table, run at maximum reasoning effort, tells a more favourable story on agentic and coding work:
- DeepSWE 1.1: Wonder Full 79.0 and Wonder 78.8, versus 74.1 for GPT-6 Astra and 74.2 for Opus 5.5
- SWE-Marathon (long-running coding): Wonder Full 57.5, against 42.5 for Astra and 46.3 for Opus 5.5
- Harvey LAB-AA (legal work): Wonder Full 13.1, well ahead of Astra at 8.6 and Opus 5.5 at 4.2
The models are not dominant everywhere. On Terminal-Bench 4.0-AA, Wonder Full's 57.1 trails both Astra (59.1) and Opus 5.5 (59.6), and it also falls short of both rivals on HealthBench Pro. Meta additionally claims Wonder beats Fable, Opus and Astra on MemoryBench, a test of continual learning and memory construction, though TechCrunch notes the claim comes from Meta's model card rather than independent testing.
VentureBeat cautions that Meta's comparison mixes internal, vendor-supplied and third-party numbers, and that some coding results were produced with different tooling, so they are less tightly controlled than a head-to-head evaluation.
Memory and Persistence as the Differentiators
Rather than leading on raw intelligence scores, Meta is pitching two capabilities aimed at long-running assistants and agents:
- Active Memory retains personal and professional context across conversations, so users do not have to re-explain themselves each session.
- Active Focus keeps an agent working toward a goal after it hits errors or obstacles, while asking for permission before taking consequential actions.
Both features underpin the Muse agent, which the Wonder models will now power. Meta is targeting use cases including software development, financial analysis, legal document preparation and insurance workflows.
Pricing, Access and the End of Open Weights
Standard Muse Wonder is available through the Meta Model API at $4 per million input tokens and $20 per million output tokens, with cached input at $0.20 per million. That matches Claude Opus 5.5's list price and undercuts GPT-6 Astra and Claude Fable 5.1, which are listed at $10 and $50. Artificial Analysis estimates Wonder's average cost per Index task at $4.52, compared with $5.98 for Opus 5.5 and $3.26 for Astra. The API supports interfaces compatible with OpenAI and Anthropic tooling, lowering switching costs for developers.
Wonder Full is a different matter. It is offered only inside Muse Code subscriptions at $50 or $200 per month, with no pay-as-you-go API price announced. Those subscriptions default to a "Contributor" mode that lets Meta use prompts, code and outputs to improve its models, although users can opt out for future submissions. The standard API mode, by contrast, does not use submitted content for training.
Most notably, neither model can be self-hosted. For a company that built its AI reputation on openly released Llama weights, that is a significant strategic shift.
Why It Matters
Muse Wonder confirms that Meta has rejoined the frontier race, and that its massive compute investments, including the Prometheus cluster, are now producing competitive models. A fourth credible closed-model vendor at the top of the leaderboards adds pricing pressure for buyers and gives developers another option for agentic coding and long-context work.
It also marks a turning point for open-source AI. Meta was long the strongest Western champion of open weights; its best models are now proprietary, leaving that role increasingly to Chinese labs and smaller Western players such as Mistral and Reflection AI. For enterprises, the practical caveats are real: Wonder Full is not on the API, nothing can run on-premises, and teams handling sensitive data will need to review Meta's service modes and retention terms carefully before adoption.
