October 12, 2026
Muse Wonder: Meta releases agent model, API priced at $4 per million input tokens
Meta released Muse Wonder on Oct 12: via the API the model costs $4 per million input tokens and $20 per million output tokens. The full version, Muse Wonder Full, is so far available only to Muse Code subscribers, and Meta has not published a per-token price for it.
Artificial Analysis
@artificialanlys
Meta released Muse Wonder and Muse Wonder Full: at maximum reasoning they scored 52 and 56 points on the Artificial Analysis Intelligence Index v4.3. Each gained noticeably over Muse Spark 1.3 (max), which scored 48. Only six weeks passed between the releases. Muse Wonder is available to users through Meta's first API, and Muse Wonder Full is so far only for Muse Code subscribers. Meta says both models are built for personal and professional agents. Congrats to @AIatMeta, @finkd and @alexandr_wang on the release! Highlights: ➤ Both models gained noticeably on the overall index. Muse Wonder Full (max) improved on Muse Spark 1.3 (max) across all ten index evaluations, gaining 8 points to reach 56. The biggest gains came in agentic evaluations: Terminal-Bench 4.0 (+24 points, from 33% to 57%), GDPval-AA v2.1 (+176 Elo, from 1681 to 1857) and AA-Briefcase (+148 Elo, from 1581 to 1729). Muse Wonder (max) gained 4 points to reach 52, rising sharply on Terminal-Bench 4.0 (from 33% to 52%) and AutomationBench-AA (from 58% to 68%). ➤ Muse Wonder Full (max) ranks in the top three in six of the ten index evaluations and is especially strong at knowledge-work agent tasks and scientific reasoning. On knowledge-work agent tasks, it narrowly leads GDP.pdf (33%), ranks second on GDPval-AA v2.1 (1857 Elo), and third on AA-Briefcase (1729 Elo) and AutomationBench-AA (71%). On scientific reasoning, it tops the CritPt table with 33% and is third on SciCode with 63%. This continues the strengths of Muse Spark 1.3 (max), which also stood out in these two areas. ➤ Both models use fewer output tokens than Muse Spark 1.3 (max). Muse Wonder (max) and Muse Wonder Full (max) each use about 54 thousand output tokens per index task, 10% fewer than Muse Spark 1.3 (max) at 60 thousand. That is less than Gemini 4 Argon (high, 62 thousand) and Claude Fable 5.1 (max w/ fallback, 78 thousand), but more than GPT-6.1 Sol (max, 38 thousand) and GPT-6 Astra (max, 27 thousand). Muse Wonder Full uses 16% fewer input tokens per task than Muse Wonder (max) (6.44 million vs 7.69 million), so it is both smarter and more token-efficient. ➤ Cost per task for Muse Wonder (max) is higher than for most models at a similar level. The premium API price ($4 per million input tokens and $20 per million output tokens) and heavy input usage (7.69 million per task) add up to $4.52 per index task. That is below Claude Fable 5.1 (max w/ fallback, $7.63), Claude Opus 5.5 (max w/ fallback, $5.98) and Claude Sonnet 5.5 (max w/ fallback, $5.46), but above GPT-6 Astra (max, $3.26), Gemini 4 Argon (high, $1.99), GPT-6.1 Sol (max, $0.72) and its predecessor Muse Spark 1.3 (max, $1.60). Meta has not yet published a per-token price for Muse Wonder Full, so its cost per task can't be calculated at this stage. ➤ Both models rose sharply on the coding agents index. Muse Wonder Full scores 68 at max level in Muse Code, level with Claude Code on Sonnet 5.5 (max, 68) and above Opus 5.5 (max, 66). Muse Wonder (max) scores 63. That is a clear gain over Muse Code on Muse Spark 1.3 (max), which scored 54. Other details of both new models: ➤ Context window: 1 million tokens each, same as Muse Spark 1.3 ➤ Price: for Muse Wonder, $4 per 1 million input tokens, $0.20 for cached input and $20 for output. No per-token price has been announced for Muse Wonder Full yet ➤ Input modalities: text, images, video and audio ➤ Output modalities: text only ➤ Availability: Muse Wonder on Meta's first API and in Muse Code from launch; Muse Wonder Full only for Muse Code users
· 5.2K views
Muse Spark 1.3 scored 48 points on the Artificial Analysis index, Muse Wonder scored 52, and Muse Wonder Full scored 56. Six weeks passed between Muse Spark 1.3 and the new models. Terminal-Bench 4.0 grew the most: Full rose from 33% to 57%, and Muse Wonder to 52%.
Coding. On the coding agents index in Muse Code, Muse Spark 1.3 scored 54, Muse Wonder 63, and Muse Wonder Full 68. Full scored the same as Claude Code on Sonnet 5.5, and higher than Opus 5.5 (66).
Cost per task. Muse Wonder costs $4.52 per index task: less than Claude Fable 5.1 ($7.63), Opus 5.5 ($5.98) and Sonnet 5.5 ($5.46), but more than GPT-6 Astra ($3.26) and Muse Spark 1.3 ($1.60). Input is the culprit: at $4 per million tokens, the task uses 7.69 million input tokens. Meta has not yet published a per-token price for Muse Wonder Full, so its cost per task can't be calculated.
Also. Context window of 1 million tokens, same as Muse Spark 1.3. New features: Active Memory keeps context between conversations, and Active Focus handles complex tasks, tries different approaches and asks clarifying questions. Muse Wonder is coming to the Muse app, subscribers first.
Next milestone: Muse Wonder comes to the Muse app, subscribers first. No date yet.