Research|META: Muse Wonder Returns Meta to the Frontier as Compute Tightens

On October 12, Meta Superintelligence Labs (MSL) released Muse Wonder and Muse Wonder Full, the models previously referred to in reports as Watermelon. These are Meta’s first frontier models since Llama 4, arriving 14 months after Alexandr Wang reorganized the company’s AI teams in August 2025. Both were trained on Meta’s Prometheus cluster in Ohio, support text, image, audio, and video inputs, and have a 1M-token context window. Neither model’s weights are publicly available.

Wonder is available through the Meta Model API and Muse Code at $4 per million input tokens and $20 per million output tokens. Wonder Full is available only to Muse Code subscribers on the $50 and $200 monthly plans, with no API access. Our channel checks indicate that Wonder has roughly 3T total parameters, making it smaller than both Astra and Fable.

We believe Meta is back among the leading frontier AI labs, and we expect it to stay there. In our view, a year ago that group consisted mainly of OpenAI and Anthropic. Google rejoined this year, followed by Meta, and a four-way contest is taking shape. The hiring spree in the summer of 2025 is paying off: backed by substantial investment in compute, the new team brought Meta back to the frontier in just over a year.

Compute remains the biggest bottleneck. According to media reports, in July Meta was preparing to rent out surplus GPUs and was in talks with Anthropic; by October, its own needs had forced it to shelve the plan. Anthropic, meanwhile, secured new capacity from other suppliers and raised usage limits when it launched Opus 5.5. We expect the shortage to ease only once Vera Rubin and TPU 8t ship in volume in 2027. We also expect Meta’s 2027 capex to rise again. On open-weight models, we maintain our September view that the gap between the leading open-weight models and the top closed models will widen again to 6 to 12 months.

Meta returns to the top tier

The independent evaluations put Wonder Full among the leading models. As of October 12, the Artificial Analysis (AA) Intelligence Index, a weighted composite of 10 evaluations, scores Opus 5.5 at 58; Wonder Full and Sonnet 5.5 at 56; Astra, Gemini 4 Argon, and Fable 5.1 at 53; and Wonder at 52. AA flags verbosity as a weakness: Wonder Full used 140 million output tokens to complete the full suite, versus a median of 88 million across the comparison group on that page. AA estimates the average cost per index task at $4.52 for Wonder, $5.98 for Opus 5.5, and $3.26 for Astra.

Meta’s model card compares models at the reasoning settings specified in its comparison. Wonder Full’s strengths are in long-horizon agent tasks and professional work. On SWE-Marathon it scores 57.5 at xhigh, against 42.5 for Astra and 46.3 for Opus 5.5; on DeepSWE it scores 79.0, almost 5 points ahead of both. On Harvey’s legal agent benchmark, Wonder Full, Astra, and Opus 5.5 score 13.1, 8.6, and 4.2, respectively. Absolute scores are low, but Wonder Full’s lead is substantial. On JobBench it scores 75.3, essentially tied with Opus.

Wonder Full trails Astra and Opus in terminal-based coding and healthcare. It scores 57.1 on Terminal-Bench 4.0-AA, 2 to 2.5 points behind Astra and Opus, and 62.3 on HealthBench Pro, about 3 points behind. On MemoryBench, Meta’s internal benchmark, Wonder scores 85.6 and Wonder Full 85.5, versus roughly 81.5 for Opus 5.5 and Astra. Memory is a key focus of this release, though these results come from Meta’s own testing.

In June 2025, Meta paid $14.3b for a 49% stake in Scale AI and brought in Wang as Chief AI Officer. Nat Friedman and Daniel Gross joined soon after. In August, Meta reorganized its AI operations into what is now MSL and recruited researchers from OpenAI, Google DeepMind, and Apple. Reported compensation packages ran to nine figures, and OpenAI CEO Sam Altman publicly cited $100m signing bonuses at the time. The hiring drew plenty of skepticism: Meta could pay to attract talent, but whether it could build a frontier lab remained an open question.

Just over a year later, MSL has delivered Wonder. Planning for Wonder, internally codenamed Watermelon, began in August 2025; the Spark line carries the codename Avocado. The team first validated its training recipe, data, and RL environments on a small model, then increased model size by a factor of 4 to 5 with each generation. Spark only began pretraining at the end of 2025, and Wonder Full now scores 56 on the independent index, tied with Sonnet 5.5 and second only to Opus 5.5. In our view, this represents faster iteration than Meta achieved in the Llama era.

We expect Meta to remain in the top tier. The team has proven itself, the Prometheus expansion and the Hyperion build-out will add compute, and Meta’s existing apps provide established distribution channels. We believe progress on the next model depends mainly on when that new capacity comes online.

Specialized models and existing users support personal assistants