Meta releases Muse Wonder and Muse Wonder Full models, scoring 52 and 56 on the Artificial Analysis Intelligence Index at roughly triple the per-task cost of Muse Spark 1.3

Meta released two models, Muse Wonder and Muse Wonder Full. Muse Wonder is available through Meta's API and Muse Code. Muse Wonder Full is limited to Muse Code users, with no token price announced. Both take text, image, video and audio input and have a 1M-token context window. On the Artificial Analysis Intelligence Index they score 52 and 56, against 48 for Muse Spark 1.3. Terminal-Bench 4.0 rose from 33 to 57 percent, and the hallucination rate fell from 33 to 23 percent. API prices for Muse Wonder rose to $4 per million input tokens and $20 per million output tokens, from $1.25 and $4.25. The cost per task on the Intelligence Index rose to $4.52 from $1.60.

Developers and enterprises evaluating coding agents now have another credible option. Gains in agentic coding are large, with a Coding Agents Index of 63 to 68. But the price is roughly triple per task, so the choice depends on whether the extra capability pays for itself in your workload. Factual knowledge is weak (AA-Omniscience 25 to 30, against 42 to 46 for rivals), which limits use for knowledge-heavy tasks. Because Wonder Full has no price and is tied to Muse Code, it is hard to plan around yet. The release shows Meta is back in the frontier conversation, but it does not change who leads.

"Closed the gap" and "breaks into the top tier" overstate it: Meta is still behind on factual knowledge and is now a fourth strong competitor, not a leader. The tripled cost per task is the part that is easy to bury. Claims about the underlying base model are the outlet's guess, and the launch framing around a personal AI agent is promotion, not evidence. Most numbers also come from Meta's own materials and third-party benchmarks, so independent real-world testing is still thin.

1. 01Whether Meta publishes API pricing and wider access for Muse Wonder Full, and at what cost per task compared with Muse Wonder
2. 02Independent coding-agent results on real repositories, and whether the Terminal-Bench 4.0 gain (33 to 57 percent) and Artificial Analysis scores hold up in third-party reruns over the next few weeks
3. 03Developer uptake and any price response from Anthropic, OpenAI or Google, plus how quickly Muse Wonder arrives in the paid Muse app and whether the weak factual knowledge draws user complaints