Meta Launches Muse Wonder Models For Coding And AI Agents
Meta says its new Muse Wonder models rival leading AI systems, with access through Muse Code and an API and plans to bring them to its Muse agent.
Meta has introduced Muse Wonder and Muse Wonder Full, two AI models it says can compete with leading systems on complex tasks. The launch extends the company’s push into more capable AI after the release of its smaller Muse Spark model. Wonder is available through the Meta Model API and Muse Code, while Wonder Full is limited to Muse Code subscribers, according to TechCrunch.
Table of Contents
What Meta is claiming
Meta describes the pair as a major step toward what it calls “personal superintelligence.” According to TechCrunch, the company’s tests put the models at benchmark scores comparable to Fable 5.1 and GPT-6 Astra. Wired reports a more specific claim for the largest Wonder model: Meta says it ranks below Claude Opus 5.5 but above Claude Fable 5.1 on the Artificial Analysis Intelligence Index, while outperforming those models and GPT-6 Astra on some specialized legal-document and software-development tests.
Those are Meta’s reported results, not a guarantee of how the models will perform in everyday use. Wired cautions that benchmarks do not always predict real-world performance. The Verge also notes that it remains unclear how the new models compare with Google’s Gemini 4 Argon.
Memory, focus and access
Meta is emphasizing two capabilities called active memory and active focus. Engadget says active memory is intended to carry relevant information between conversations, while active focus is meant to help an agent keep working toward a goal when a task becomes difficult. TechCrunch reports that, in Meta’s testing, Muse Wonder scored above Fable, Opus and Astra on MemoryBench, a test of memory and continual learning.
Meta plans to use the models in its Muse personal agent, beginning with subscribers, according to Engadget and The Verge. The Verge says Meta has not provided a timeline or specifics for that rollout. For now, The Verge reports that both models are available to Muse Code subscribers and that Wonder can also be accessed through Meta’s API.
Engadget lists API pricing for Wonder at $4 per million input tokens and $20 per million output tokens. It says Wonder Full is available through Muse Code’s $50-per-month power usage and $200-per-month maximum usage tiers; TechCrunch says Full will not be offered through the API.
What developers can watch
For people building software with AI, the immediate change is another model option in Muse Code and, for Wonder, through an API. Meta claims the models perform strongly on coding tasks, but the reported benchmarks leave open how they will fare on a particular team’s work.
Wired says Meta has also trained Wonder to refuse dangerous requests and reports that the company cites favorable safety-test results. The outlet notes that Meta has not disclosed whether agents powered by Wonder misbehaved during testing or how it addressed that specific risk.
Sources
This story was compiled by AI from the reports below. Read the originals for the full details.