Meta's New Muse Wonder Finally Puts Zuckerberg in the Frontier AI Race
Background, discussion, and references
Market impact
The direct channel is pricing pressure on frontier inference: Meta is matching Claude Opus 5.5 at $4/$20 per million tokens and undercutting GPT-6 Astra's $10/$50, which affects the API revenue economics of competing labs rather than any crypto asset directly. The read-through to crypto markets runs through sentiment — Meta's positioning as a frontier lab feeds the broader AI capex and risk-appetite narrative that typically moves high-beta technology-linked tokens alongside other risk assets.
Background
Meta formed Meta Superintelligence Labs in 2025 after its open-source Llama 4 models stumbled in independent tests, following a roughly $14 billion deal for Scale AI that put founder Alexandr Wang in charge of the unit. The group's first model, Muse Spark, launched earlier in 2026 as a small, fast, closed model. Muse Wonder arrives weeks after OpenAI shipped GPT-6 Astra on September 3 and Anthropic followed on September 22 with Claude Opus 5.5.
References
Tags
#Meta#Muse Wonder#frontier models#benchmarks#AI competition#Meta Superintelligence Labs
Meta released Muse Wonder and Muse Wonder Full on Monday, the first frontier-class models from Meta Superintelligence Labs, both with a 1 million token context window. The API version, Muse Wonder, is priced at $4 per million input tokens and $20 per million output tokens, while the larger Muse Wonder Full is reserved for paying Muse Code subscribers.
This is the first time Meta has been willing to put its own models head-to-head with the best from OpenAI and Anthropic. On Meta's own charts Wonder Full leads GPT-6 Astra and Claude Opus 5.5 on long-horizon coding, legal agent work and memory, but the third-party Artificial Analysis ranking still places Opus 5.5 first at 58 versus Wonder Full's 56.
Muse Wonder is live now through the Meta Model API and Muse Code, Meta's coding agent. Meta presents two features — Active Memory, which carries information from one conversation into the next, and Active Focus, which keeps the model working toward a hard goal and checking in before doing anything important — and it acknowledges that the SWE-Marathon coding row is not a controlled comparison and its model ran at maximum effort.
rss · Decrypt · · Single source