Meta debuts Muse Wonder and Muse Wonder Full with 1 million token context

Meta has released two new large language models, Muse Wonder and Muse Wonder Full, each built with a 1 million token context window, the stretch of text a model can read in a single request. Meta describes the pair as its first frontier-class models, a term the company uses to mean they are meant to compete with the strongest systems from other AI labs. Both are built as agent models, designed to carry out multi-step tasks rather than answer a single question and stop.

Muse Wonder, the smaller of the two, is priced at $4 and $20 per million tokens on the Meta Model API, the interface developers use to call Meta’s models directly in their own code. The available report does not say which of those two rates covers text sent into the model and which covers text the model generates back, a split most AI providers list separately. Until Meta publishes a full pricing page, that detail stays open.

Muse Wonder Full, the larger model, is not sold through the same pay-per-token API. Instead, it is bundled into Meta’s Muse Code plans, a subscription-style access route rather than metered billing. According to the report, Muse Wonder Full leads DeepSWE and SWE-Marathon, two benchmarks that test how well AI agents handle software engineering work such as fixing bugs or completing coding tasks end to end. No specific scores from those benchmark runs have been published alongside the announcement.

What’s confirmed and what isn’t

The headline claim, that Meta now has frontier-class models with competitive reasoning ability, comes from Meta itself rather than from an outside test. Aside from the 1 million token context window and the Muse Wonder pricing, few other technical specifics are available: there is no confirmed parameter count, no description of training data, and no stated release date for Muse Wonder Full’s own pricing or wider rollout. Independent developers have not yet had time to run their own checks against DeepSWE and SWE-Marathon or to confirm Meta’s reported results.

That gap matters because benchmark leadership claims made by a model’s own creator can differ from how a model performs once outside users put it to work on their own codebases and tasks.

What to watch next is whether Meta publishes pricing and access details for Muse Wonder Full itself, and whether independent developers post their own DeepSWE and SWE-Marathon results once they get hands-on access to both models.