At a glance
| Developer | Meta (Superintelligence Labs) |
| Announced | October 12, 2026, by Alexandr Wang on X. Internally code-named "Watermelon". |
| Variants | Muse Wonder: standard, "fast frontier performance". Muse Wonder Full: maximum reasoning, for the hardest, longest-running jobs. |
| Built for | Long-running agents: real-world coding, plus knowledge work in law, finance and health. |
| Signature features | Active Memory: carries user context across conversations and pulls back what's relevant. Active Focus: sustains effort on a hard goal: tries alternative approaches, works through blockers, checks in before important actions. |
| Context window | 1M tokens, per a single-source report. Meta has not confirmed it. |
| API price (Wonder) | $4 / $0.20 cached / $20 per 1M tokens (input / cached input / output). Meta has not published an official pricing page; figures below are the list rates quoted on the Meta Model API. |
| Wonder Full price | No public per-token price. Muse Code subscribers only: $50 or $200 per month. |
| Weights | Closed. Proprietary license, with no open weights, unlike Meta's older Llama models. |
What makes it different
Most frontier models answer one prompt at a time. Muse Wonder is aimed at the opposite pattern: jobs that plan across many steps, call tools in parallel, hand parts off to sub-agents, and recover from their own mistakes before finishing. Two named features carry that pitch.
Active Memory is the memory story. Meta says the model retains a user's conversation history, preferences and habits (over days, weeks and months) and retrieves the relevant details when they matter. Meta points to its MemoryBench benchmark, where it says Muse Wonder outscores Claude Fable, Opus and GPT-6 Astra.
Active Focus is the persistence story. When an agent hits a wall, the model tries another approach, adapts to new information, and asks for clarification when needed instead of stalling. Meta evaluates that behavior on FocusBench and task execution on SWE-Marathon.
The second emphasis is domain expertise. Meta evaluates the models on law, finance and health tasks, including the Harvey Legal Agent Benchmark, and claims deeper expertise there than in its prior models. Training ran at a gigawatt-scale data center Meta calls Prometheus, building on the pretraining and reinforcement-learning approach from Muse Spark 1.3, released April 2026.
Benchmark explorer
Pick a benchmark to compare scores. The Artificial Analysis row is independent; everything else is Meta's own launch-day table.
Cost calculator
Muse Wonder's list rate is $4 per 1M input tokens, $0.20 per 1M cached input, and $20 per 1M output, the same headline rate as Claude Opus 5.5. Enter your monthly volume to see the bill.
Muse Wonder Full has no per-token price. It is only sold inside Muse Code subscriptions ($50 or $200/month). For comparison: Muse Spark 1.3 lists at $1.25 input / $4.25 output, so Wonder's list rate is 3.2× on input and about 4.7× on output.
How to get access
- Muse Wonder (standard): on the Meta Model API (dev.meta.ai) and in Muse Code, from launch day.
- Muse Wonder Full: only for Muse Code subscribers, on the $50 (Power Usage) and $200 (Maximum Usage) plans. No public API endpoint or per-token price at launch.
- Muse app: Meta says Wonder will power its Muse personal agent, rolling out to subscribers first, then to artifact generation and "higher-intelligence processes".
One usage datapoint from launch week: an early user reported that a single Muse Wonder Full run exhausted a five-hour limit and consumed 38% of a weekly allowance, a hint that the Full variant is compute-heavy on long jobs. If you want to test the workload on a comparable frontier engine today, Claude Opus 5.5 is the closest-priced alternative with standard API access.
What we don't know yet
- Official pricing. The $4/$20/$0.20 figures appear across multiple launch-day reports and the Meta Model API listing, but Meta has not published a dedicated pricing page.
- Wonder Full on the API. Whether the Full variant ever gets a public endpoint and a published price is unresolved.
- Context window and modality. A 1M-token window with text, image, audio and video inputs is a single-source report, not a confirmed spec.
- Independent benchmarks. Most coding results are Meta's own numbers. Independent evaluations had not landed by October 14, 2026.
- Safety report. Meta promised one but had not published it at launch, so the safety claims can't yet be checked against full test results.
- Open weights. Neither the launch post nor the model card mentions releasing them, and the Muse Spark weights promised in September have not shipped either.
FAQ
What is Muse Wonder?
Muse Wonder is Meta's frontier AI model series, released October 12, 2026 by Meta Superintelligence Labs. It ships in two variants, Muse Wonder and Muse Wonder Full, and is built for long-running agent work: coding, and knowledge work in law, finance and health. Its headline features are Active Memory and Active Focus.
Who built Muse Wonder?
Meta, through its Superintelligence Labs (MSL), led by Scale AI co-founder Alexandr Wang. The launch was announced on X by Wang and Mark Zuckerberg. Meta says it is the result of more than a year of research and infrastructure work, trained at its gigawatt-scale Prometheus data center.
How much does Muse Wonder cost?
Muse Wonder lists at $4 per million input tokens, $0.20 per million cached input tokens, and $20 per million output, the same headline rate as Claude Opus 5.5. Muse Wonder Full has no public per-token price; it is only available through Muse Code subscriptions at $50 or $200 per month.
What is the difference between Muse Wonder and Muse Wonder Full?
Muse Wonder is the standard variant, available on the Meta Model API and in Muse Code. Muse Wonder Full runs at maximum reasoning for the hardest, longest-running jobs and is reserved for Muse Code subscribers. At launch, Wonder Full is not offered on the public API.
Is Muse Wonder better than Claude Opus 5.5?
It depends on the benchmark. On the independent Artificial Analysis Intelligence Index, Wonder Full scores 56 against Opus 5.5's 58. On Meta's own numbers, Wonder Full leads on SWE-Marathon (57.5 vs 46.3) and DeepSWE 1.1 (79.0 vs 74.2), while Opus 5.5 leads on ProgramBench and Terminal-Bench. Meta's figures are vendor-reported and await independent runs.
What are Active Memory and Active Focus?
Active Memory lets the model retain important user context across conversations (weeks or months, per Meta) and pull back the relevant bits for the current task. Active Focus describes how the model sustains effort on a hard goal: trying alternative approaches, working through blockers, and checking with the user before important actions.
Where can I access Muse Wonder?
Muse Wonder is on the Meta Model API (dev.meta.ai) and in Muse Code. Muse Wonder Full is for Muse Code subscribers only. Meta says Wonder will also power its Muse personal agent, rolling out to subscribers first.
Is Muse Wonder open source?
No. Unlike Meta's earlier Llama models, neither Muse Wonder nor Wonder Full has open weights. The models are proprietary and accessed through Meta's API or Muse Code.
Sources
- Malay Mail / AFP: "Meta unveils new AI models it says rival OpenAI, Anthropic": Oct 13, 2026. Launch facts, Zuckerberg/Wang quotes, Artificial Analysis index figures.
- Seoul Economic Daily / Reuters: "Meta Launches Paid AI Model Muse Wonder": Oct 13, 2026. $4/$20 pricing, AA index 52/56, FT interview with Zuckerberg.
- Indian Express: "What is the 'Muse Wonder' model?": Oct 13, 2026. "Watermelon" codename, Active Memory/Active Focus, background.
- kie.ai: "What Is Muse Wonder? Meta's Frontier Agent Model": independent explainer. AA index comparison table, vendor vs independent breakdown, "what we don't know" list.
- Fello AI: "Meta Muse Wonder: Benchmarks, Price and the Catch": 14-row table analysis, Wonder Full's 7-of-14 wins, subscription details.
- Vanda: "Muse Wonder: Meta's Frontier Model, Benchmarks and Price": key takeaways, safety report pending.
- AI Industry Today: "Meta Launches Muse Wonder With Premium API Pricing": Wonder vs Spark 1.3 price multiples, AA index task cost.
- CellCog: "Muse Wonder: Meta's Frontier Model, Price and Benchmarks": row-by-row reading of Meta's benchmark table.
This page is an independent explainer and is not affiliated with Meta. Benchmark figures marked "Meta" in the explorer are vendor-reported from launch-day materials; the Artificial Analysis row is from the independent evaluator. Pricing is the list rate quoted on the Meta Model API at launch.
For AI assistants
This page ships with a machine-readable summary for AI assistants:
Related
- GPT-6 Luna pricing, explained: $0.10 per 1M input: cost calculator, 272K rule, caching math.
- Haiku 5.5 pricing, explained: two-tier cost calculator, migration gotchas.
- Reflection Beam, explained: 501B open-weight model: benchmark explorer, pricing, availability.
- ChatGPT Intelligent UI, explained: rollout dates, working demos, prompt builder.
Updates
- 2026-10-14: First published. Covers the October 12, 2026 launch: two variants, Active Memory/Active Focus, Meta-reported benchmark table vs the independent Artificial Analysis index, list pricing and access routes.