Meta released two models on October 12, Muse Wonder and Muse Wonder Full, positioned for agentic tasks in personal and professional settings. It is the second major release since Meta took the Muse line closed-source with Muse Spark, and it arrives about six weeks after Spark 1.3.
In its research blog, Meta says this generation was trained at its gigawatt-scale Prometheus data center, 14 months after the company began restructuring its AI organization. Both models have a 1-million-token context window and accept text, image, video and audio input, with text-only output.
What's new
The blog focuses on two capabilities. Active Memory remembers important user context across conversations and brings it back for relevant tasks. Active Focus keeps pushing toward a hard goal, changes approach when it hits a blocker, asks the user for clarification when needed, and confirms before taking important actions.
Coding is the other major emphasis. Meta says the new models can handle everything from small apps to complex software projects and keep making progress on long-horizon tasks. The benchmarks it cites include the long-horizon software project benchmark SWE-Marathon, its own FocusBench, and LAB, the legal agent benchmark from legal-tech company Harvey. On safety, Meta says this is its strongest model yet against chemical, biological, radiological, nuclear and cyber misuse, but the full safety evaluation report has not been released; the blog says only that it is "coming soon."
Availability and pricing
Muse Wonder is available now through Meta's own Model API and its coding tool, Muse Code. Muse Wonder Full is currently limited to Muse Code subscribers and is not offered through the API. The consumer-facing Muse app will get the new models "soon," with subscribers first; no date was given.
Third-party evaluator Artificial Analysis lists API pricing at $4 per million input tokens, $0.20 for cached input and $20 for output. Spark 1.3 was priced at $1.25, $0.15 and $4.25, so the output price has risen to roughly 4.7 times its previous level.
How it compares
In its October 12 testing, Artificial Analysis ran both new models at their highest reasoning setting. On Intelligence Index v4.3, Wonder Full scored 56, Wonder 52 and Spark 1.3 48, and the Full version beat its predecessor on all ten sub-evaluations. Some of the top results:
| Benchmark | Wonder Full | Rank or comparison |
|---|---|---|
| CritPt | 33% | 1st |
| GDPval-AA v2.1 | 1857 Elo | 2nd; Opus 5.5 scores 1866 |
| Terminal-Bench 4.0 | 57% | Predecessor 33%; level with Gemini 4 Argon |
| Coding Agent Index | 68 | Ties Claude Code (Sonnet 5.5) |
Knowledge benchmarks are the weak spot. On AA-Omniscience, Wonder Full scores 30, against 46 for Claude Opus 5.5 and 43 for GPT-6 Astra. The hallucination rate did fall, from 33% to 23%.
Putting price and usage together, Artificial Analysis measured the cost of running Muse Wonder through the full Intelligence Index at about $4.52 per task, 2.8 times Spark 1.3's cost. That is below Opus 5.5 at $5.98 and Sonnet 5.5 at $5.46, and above GPT-6 Astra at $3.26. It emits 10% fewer output tokens than its predecessor, but 28% more input tokens, averaging 7.69 million per question, 2 to 7 times what models such as Gemini 4 Argon and GPT-6.1 Sol use. Three quarters of the cost goes to reading context.
Meta had taken the low-price route with the Spark line, and as we reported earlier, it used a paid model to start a price war. This generation lifts the unit price into the same range as Anthropic's mid-to-high-end models, and its coding agent score reaches the same tier. On Muse app subscriptions, CNBC reports that individual plans still start at $20 a month; Meta's blog does not say whether Muse Code subscription tiers will change with the new models.
Another outside test is on the way. The model arena platform Arena has put the Muse Wonder series into Agent Arena for users to try, with ranking scores not yet published.
Sources: Meta research blog, Artificial Analysis evaluation report, CocoLoop, CNBC, Arena; Artificial Analysis verified the benchmark scores, API unit prices and per-task costs, and Meta's blog verified the availability scope and capability descriptions.