Meta launches Muse Wonder: what it is, what it costs, and why people are arguing about it

On Monday, 12 October, Meta introduced Muse Wonder, a new family of models it calls its most capable yet. There are two of them: Muse Wonder and the bigger Muse Wonder Full. Meta pitches both at what it calls personal and professional agentic work, meaning long, multi-step jobs done on your behalf.

Wonder vs. Wonder Full

• Muse Wonder (model ID muse-wonder-1.0): available now through the Meta Model API and Muse Code. Meta's developer page lists a 1M-token context window and API pricing of $4 per million input tokens, $0.20 cached input and $20 output.
• Muse Wonder Full (muse-wonder-full-1.0): available now only to Muse Code subscribers, on the $50/month Power Usage and $200/month Maximum Usage plans. Meta has not published API pricing for it.
• Muse Wonder will also power parts of Muse, Meta's personal agent app, rolling out to subscribers first.

Meta's newsroom post highlights two features: Active Memory, which carries context across conversations, and Active Focus, which keeps the model working on hard tasks, trying other approaches and checking in before important actions. Meta says the models were released under its Meta Superintelligence Scaling Framework and that a full safety report is coming.

What the independent benchmarks say

Meta's own charts compare the models with other frontier systems, but outside numbers matter more. Artificial Analysis puts Muse Wonder at 52 and Muse Wonder Full at 56 on its Intelligence Index v4.3 (max reasoning effort), up from 48 for Muse Spark 1.3, which shipped about six weeks earlier. On its Coding Agents Index, Wonder Full in Muse Code scored 68, level with Claude Code running Sonnet 5.5 (68) and just above Opus 5.5 (66).

Artificial Analysis also flags cost. Muse Wonder used a lot of input tokens per task, so it cost $4.52 per Intelligence Index task. That's below several Claude models but above GPT-6 Astra ($3.26) and GPT-6.1 Sol ($0.72). Vals AI says Wonder Full (Max) debuts at #5 on its Vals Index, ahead of every OpenAI model it tracks.

How Meta got here

According to The New York Times, the models were codenamed Watermelon and come from the AI division led by Alexandr Wang. Meta paused pretraining in July, which pushed the launch to October. The Muse agent app launched in September. Quartz gives the date as 8 September and cites Sensor Tower: 5 million downloads in 22 days.

Hype vs. doubt

The first reactions on X split fast. Theo of t3.gg declared Meta a frontier lab. Ethan Mollick, who had early access, called it a good model in limited testing and pointed to how quickly Meta had improved since April. Others were cautious. Bindu Reddy suspected the model was tuned to benchmarks, and Lisan al Gaib still put Meta a few months behind.

Launch day also had a practical hiccup. Matthew Berman posted that he "almost immediately ran out of quota", and Alexandr Wang later wrote that Meta was "working to fix the limits issue" because limits weren't behaving as intended.