Meta launches Muse Wonder and Wonder Full models, says they rival OpenAI and Anthropic flagships

Meta is picking up the pace in the AI race. The company has launched two new models, Muse Wonder and Muse Wonder Full, opening them first to Muse Code subscribers while planning to wire Wonder into its Muse AI agent — and claiming the new models can go toe-to-toe with OpenAI's and Anthropic's flagships in internal testing. The thread ties together a model launch, subscription monetization, API access, and the agent ecosystem.

Two new models land first on subscriptions and API

Meta announced Muse Wonder and Muse Wonder Full, both available to Muse Code subscribers starting today. Meta said Wonder can also be accessed through the Meta Model API. That means Meta isn't keeping the new models inside its own consumer products alone; it is serving subscribers and developer call scenarios at the same time. For a company playing catch-up in the AI race, this dual-track distribution both validates model capability quickly and accumulates usage feedback through the API. Muse Code subscribers becoming the first users shows Meta chose to serve a high-engagement, more professional audience first before gradually widening coverage. ^[6]

Matching the flagships: the gap between internal benchmarks and outside validation

Meta says the new models can compete with leading models from OpenAI and Anthropic, citing GPT-6 Astra, Claude Opus 5.5, and Claude Fable 5.1 as comparison points from internal testing. Internal benchmarks show the team's confidence in its models, but third-party evaluations have yet to appear; Meta also did not disclose test methods, task types, or win rates. More notable is that Meta says it is not yet clear how the new models compare with Google's Gemini 4 Argon, whose capabilities are still taking shape. Benchmarking some rivals while leaving another blank reflects a competition that has entered a multipolar phase — and reminds readers that claimed competitiveness needs independent evaluation and real workloads to verify. ^[6]

Next steps for the Muse agent and specialized models

Meta plans to eventually power its Muse AI agent with Wonder, starting with subscribers, and then to develop specialized models so that everyone using Muse can get the relevant capabilities. It has not shared specifics or a timeline, though. Embedding model capability in an agent shifts the competitive focus from single-shot question answering to sustained task execution: models need to understand context, call tools, and complete tasks reliably. The subscribers-first, everyone-later rhythm suggests Meta wants to polish the product within a controllable scope before scaling up. Meta previously launched Muse Spark and the image model Muse Image, so the Muse line is stretching from the model layer to the agent layer — an attempt at a closed loop running from underlying models to the user's front door. ^[6]

A challenger's spending and the competitive landscape

These launches are part of Meta's effort to re-establish itself as a major AI power. The company has poured billions of dollars into talent and infrastructure and keeps expanding its AI lineup, including the image model Muse Image and its push to bring useful AI agents to the masses through Muse. Against rivals such as OpenAI, Anthropic, and Google, Meta's edge lies in its huge user base and distribution muscle, but the models themselves still need to prove out. Running subscriptions and an API in parallel lets it serve both individual professional users and developers; if the agent truly ships, it could change how users interact with AI. ^[6]

What to watch next

Next, watch for third-party evaluations of Muse Wonder and Wonder Full, the availability and developer experience of the Meta Model API, how Wonder actually performs when it powers the Muse agent, and the timeline for the 'specialized models' reaching all Muse users. By binding its new models to an agent, Meta signals that it does not just want a spot on model leaderboards — it wants to own the entry point to users' everyday tasks. The AI race is therefore shifting from raw model parameters or benchmark scores to a compound contest of subscriptions, APIs, agents, and distribution ecosystems; whoever can reliably turn model capability into a dependable product experience is the one likely to hold ground in this round. ^[6]