Meta’s Muse Wonder gains in coding, with higher API costs
Illustration, not documentary evidence of the event.
Meta has released Muse Wonder and Muse Wonder Full, with stronger coding and autonomous-task benchmark results than their predecessor, Muse Spark 1.3. According to the-decoder.com, Full scores 56 on the Artificial Analysis Intelligence Index, compared with Wonder’s 52 and Spark 1.3’s 48. Access differs: Wonder is available through Meta’s API and Muse Code, while Full is currently limited to Muse Code users.
The improvement comes with a higher price for the API-accessible model. Artificial Analysis puts Muse Wonder’s cost at $4.52 per Intelligence Index task, versus $1.60 for Spark 1.3. Wonder’s token rates are $4 per million input tokens and $20 per million output tokens. The report says token pricing for Full is not yet available, so its stronger scores do not establish which model offers better value through an API.
For developers choosing a coding agent, the results justify a trial on their own work. Full scores 68 on the Coding Agents Index, tying Claude Code with Sonnet 5.5; Wonder scores 63, up from its predecessor’s 54. A useful comparison would give Wonder and the incumbent model the same representative tasks, then record whether each passes the project’s checks, how much human correction it needs and its total cost per accepted result. Test Full separately in Muse Code if that is a usable route for the team. Its benchmark lead should not be treated as an API capability available today.
The more important distinction is between coding performance and factual reliability. Full scores 30 on the AA-Omniscience knowledge test and Wonder scores 25, below the competing models listed in the report. Full’s reported hallucination rate falls from its predecessor’s 33 percent to 23 percent, but that test result does not establish an error rate for a particular workflow. Teams evaluating research-heavy tasks should check answers against supplied reference material as well as measuring task completion. The report also says finance and legal benchmark performance stayed around predecessor levels, with declines of up to two points, giving readers little basis to assume a broad improvement across those tasks.
Meta’s personal-agent pitch adds another evaluation question. The company describes “Active Memory” as retaining information across conversations and “Active Focus” as persisting through difficult tasks, trying alternatives and asking clarifying questions. Before using those features with calendars, messages or documents, check what information is retained, how it can be corrected or removed, and which actions require approval. These are deployment criteria to verify, not shortcomings established by the report. For a team considering the upgrade, the decisive evidence would be more accepted work at an acceptable cost, with memory and permissions that fit its workflow.