简体中文版 (machine-assisted translation, not yet reviewed by a native speaker)
Meta Muse Wonder Review: Strong Agent Scores, High Price
Meta's new Muse Wonder model scores 52 on the independent Artificial Analysis Intelligence Index, ahead of Meta's own Muse Spark 1.3 (48) and Anthropic's budget Claude Haiku 5.5 (43), but it is priced at $4 per million input tokens and $20 per million output tokens, about 40 times Haiku 5.5. Meta released it on 12 October 2026, and most of its detailed claims are still Meta's own.
Key takeaways
- Muse Wonder ranks 9th on Artificial Analysis's index with a score of 52, according to Artificial Analysis. Meta's larger Muse Wonder Full is reported at 56 by a third party.
- API pricing is $4.00 input and $20.00 output per million tokens, per Artificial Analysis. It used 130 million output tokens on the index run versus a median of 85 million, which raises real-world cost.
- Meta has not given country-by-country availability, and its safety report is "available soon", not yet published.
- Meta's own agent benchmarks (FocusBench, MemoryBench) have no independent results yet.
What Meta released
Meta announced two models on 12 October: Muse Wonder and Muse Wonder Full. According to Meta, Muse Wonder is available now through the Meta Model API and the Muse Code coding tool, while Muse Wonder Full is limited to Muse Code subscribers. Meta says Muse Wonder will later power its Muse assistant, subscribers first, with no date given.
Meta describes two headline features. "Active Memory" carries relevant details across conversations. "Active Focus" keeps a model working on a hard goal, retrying other approaches and checking in before important actions. Meta's target uses are law, finance, health and coding.
Meta's model card lists a context window of 1 million tokens, text, image, video and audio input, and text-only output. It does not say whether the weights will be public; it says only that they are protected by a weight-security programme.
What independent testing shows
Artificial Analysis, an independent benchmarking firm, lists Muse Wonder at 52 on its Intelligence Index, ranked 9th of the models it tracks. It measured about 88 output tokens per second and a 19-second wait for the first token at its maximum setting, against a median of about 5.5 seconds for comparable models. It calls the model "somewhat verbose" and "somewhat expensive" for its tier.
Cost per evaluation task came to $4.52 on its index, per Artificial Analysis. A third-party write-up by AgentOne Research, whose author discloses making a competing product, reports that this is cheaper than Claude Opus 5.5 ($5.98) but dearer than GPT-6 Astra ($3.26). Treat those comparison figures as reported, not confirmed.
| Model | Input / output per 1M tokens | Artificial Analysis Index | Context |
|---|---|---|---|
| Muse Wonder (Meta) | $4.00 / $20.00 | 52 | 1M |
| Muse Spark 1.3 (Meta) | $1.25 / $4.25 | 48 | 1M |
| Claude Haiku 5.5 (Anthropic) | $0.10 / $0.50 (prompts up to 100K tokens) | 43 | 1M |
| Mistral Large 4 preview (Mistral) | $1.36 / $4.18 | not compared here | reported ~524K |
Prices are list prices as of 13 October 2026. The Mistral row is reported by Trending Topics, not independently confirmed. Index scores come from different Artificial Analysis pages and settings, so read them as indicative.
Strengths and weaknesses
Strengths: according to AgentOne Research, the largest gains over Muse Spark 1.3 are in agent-style tasks, and the model accepts text, image, video and audio input. The same write-up relays Meta's tables showing strong coding and finance-agent results, but Meta's launch text gives no scores, so treat them as unconfirmed.
Weaknesses: price, verbosity and latency. Meta's model card also lists known limits: weaker visual and spatial reasoning, especially in 3D, multi-step reasoning errors in new situations, uneven performance outside well-supported languages, and a bar on use by anyone under 18.
Best fit: teams that need a closed, high-scoring agent model and can absorb the cost. Better to wait: anyone who needs published safety results, a clear regional availability statement, or independent checks of the memory and focus claims.
The global picture
In the US, the launch follows Meta's July 2026 Muse Spark 1.1 API, which was a public preview open only to US developers, according to Implicator.ai. Meta has not said whether Muse Wonder keeps that limit. One reviewer reports API access depends on country; we could not confirm it from Meta.
In Europe, the model card names Meta Platforms Ireland Limited among the providers, but says nothing about EU availability or data residency. Mistral, the French lab, priced its Large 4 preview far lower, and Trending Topics reports it plans to release open weights at the end of October with a licence not yet named.
On price, Anthropic's own page lists Haiku 5.5 at $0.10 input and $0.50 output per million tokens, and says it suits narrow, high-volume work, with Sonnet 5.5 and Opus 5.5 recommended for complex agentic coding. VentureBeat, via Trending Topics, notes that Chinese open-weight models are often cheaper still. We have not verified that claim here.
What to watch next
- Meta's full safety evaluation report, promised as "soon".
- Independent scores for FocusBench, MemoryBench and SWE-Marathon.
- An API date and price for Muse Wonder Full.
- Mistral's open-weights release, due at the end of October (reported).
Our take
Muse Wonder is a credible step up for Meta, and the independent score backs that up. But at $20 per million output tokens and with heavy token use, it competes on capability, not on cost. Until the safety report and regional terms arrive, we would trial it, not commit to it.
FAQ
What is Meta Muse Wonder? It is an agent-focused model from Meta, released on 12 October 2026, aimed at personal and professional tasks such as law, finance, health and coding.
How much does Muse Wonder cost? Artificial Analysis lists $4.00 per million input tokens and $20.00 per million output tokens. Meta's launch post gives no price.
Is Muse Wonder open source? Meta has not said whether its weights will be released.
Is it available in India or Europe? Meta has not stated country-level availability.
Sources
- Meta, "Introducing Muse Wonder and Muse Wonder Full", 12 October 2026. https://about.fb.com/news/2026/10/introducing-muse-wonder-models/
- Meta Research, "Muse Wonder: A Frontier Model for Personal and Professional AI Agents", 12 October 2026. https://research.meta.ai/blog/muse-wonder
- Meta, Muse Wonder model card. https://research.meta.ai/static/muse-wonder-model-card
- Artificial Analysis, Muse Wonder model page. https://artificialanalysis.ai/models/muse-wonder
- Artificial Analysis, Claude Haiku 5.5 vs Muse Spark 1.3. https://artificialanalysis.ai/models/comparisons/claude-haiku-5-5-vs-muse-spark-1-3
- Anthropic, Claude Haiku 5.5, 7 October 2026. https://www.anthropic.com/claude-haiku-5-5
- AgentOne Research, "Meta Muse Wonder: Benchmarks, Pricing, API" (competitor disclosure). https://www.agent-one.dev/blog/muse-wonder
- Implicator.ai, Meta Model API and Muse Spark 1.1, July 2026. https://www.implicator.ai/meta-prices-its-first-paid-ai-model-api-at-4-25-per-million-output-tokens.md
- Trending Topics, Mistral Large 4 and Artificial Analysis. https://www.trendingtopics.eu/mistral-large-4-artificial-analysis/
- Trending Topics, Claude Haiku 5.5 (citing VentureBeat). https://www.trendingtopics.eu/claude-haiku-5-5/
AI Worldwire is independent and has no commercial relationship with the companies covered, unless stated. Research and drafting are assisted by AI; every article is reviewed and approved by a human editor before publication. Subscribe to the AI Worldwire newsletter. Last updated: 13 October 2026.