Meta Muse Wonder launched — same $4 API price as Claude Opus 5.5, six points lower on an independent index
Meta released two models on October 12, 2026: Muse Wonder and Muse Wonder Full, its first frontier-class models since Alexandr Wang restructured Meta's AI organization in August 2025. Muse Wonder costs $4 per million input tokens and $20 per million output tokens on the Meta Model API, identical to Claude Opus 5.5. Muse Wonder Full is not sold through the API and is limited to Muse Code plans at $50 and $200 a month. Meta claims state-of-the-art results on personal and professional agent work, but on the independent Artificial Analysis Intelligence Index Wonder Full scores 56 and Wonder 52, versus 58 for Opus 5.5.
The three lines
- Launch — Muse Wonder and Muse Wonder Full on October 12, 2026; multimodal input, 1 million-token context, closed weights
- Price — Wonder: $4 in / $20 out per million tokens, same as Opus 5.5. Wonder Full: no API, Muse Code $50 and $200 plans only
- Scores — independent index: Wonder Full 56, Wonder 52, Opus 5.5 58. On Meta's own 14-row chart Wonder Full wins 7, two on Meta-built benchmarks
Key questions
- Meta Muse Wonder API price
- **$4 per million input tokens and $20 per million output tokens for Muse Wonder. Wonder Full has no API price.** | Model | Input | Cached | Output | Where | |---|---|---|---|---| | Muse Wonder | $4.00 | $0.20 | $20.00 | Meta Model API, OpenRouter, Muse Code | | Muse Wonder Full | — | — | — | Muse Code $50 / $200 plans | | Claude Opus 5.5 | $4.00 | $0.20 | $20.00 | Anthropic API | | GPT-6 Astra | $10.00 | $1.00 | $50.00 | OpenAI API | | Muse Spark 1.3 | $1.25 | $0.15 | $4.25 | Meta Model API |
- Muse Wonder benchmarks vs Claude Opus 5.5
- **Wonder Full leads on long coding and legal-agent tasks in Meta's chart; Opus 5.5 leads on the independent index.** | Benchmark | Wonder Full | Wonder | Opus 5.5 | |---|---|---|---| | Artificial Analysis Index | 56 | 52 | 58 | | SWE-Marathon (Meta chart) | 57.5 | — | 46.3 | | Harvey LAB-AA (legal agents) | 13.1 | — | 4.2 | | ProgramBench | 59.0 | — | 70.5 | | MemoryBench (Meta internal) | 85.5 | 85.6 | — |
- How to use Muse Wonder Full
- **Only through paid Muse Code plans; it is not on the API.** | Route | Models | Note | |---|---|---| | Muse Code Power ($50/mo) | Wonder, Wonder Full | As reported | | Muse Code Maximum ($200/mo) | Wonder, Wonder Full | As reported | | Meta Model API | Wonder only | Pay as you go | | Muse app subscription | Wonder (coming) | Limited testing, not in every country | | Muse app free | Not yet | No date |
Meta has entered the top-model race for the first time. On October 12, 2026, Meta Superintelligence Labs released Muse Wonder and Muse Wonder Full. They are the first frontier-class models since Scale AI founder Alexandr Wang reorganized Meta's AI work in August 2025, a 14-month rebuild after the company's earlier Llama models drew a weak reception. Until now Meta had shipped the lightweight Muse Spark and a personal-agent app called Muse, while avoiding head-to-head releases against OpenAI and Anthropic. Meta's blog called the launch "a major leap toward personal superintelligence."
Line the numbers up, though, and the picture narrows. The smaller model Meta sells through its API costs exactly what Claude Opus 5.5 costs but scores six points lower on the leading independent index, and the larger model is not sold through the API at all.
1. What launched
| Item | Muse Wonder | Muse Wonder Full |
|---|---|---|
| Size | Smaller | Larger |
| Input | Text, images, video, audio | Text, images, video, audio |
| Output | Text | Text |
| Context window | 1 million tokens | 1 million tokens |
| Access | Meta Model API, OpenRouter, Muse Code | Paid Muse Code plans only |
| Open weights | No | No |
| Trained at | "Prometheus" data center, central Ohio | Same |
Meta's headline feature is "active memory," the ability to retain and combine what a user shares over time. Research scientist Kavya Kopparapu described it in the launch video as making sure the model "can remember what's important to you over days, weeks, months, and beyond." Both models will power the paid tier of the Muse agent app first; free users have no date.
2. Price and score side by side
| Model | Artificial Analysis Index | Input / 1M tokens | Output / 1M tokens | Public API |
|---|---|---|---|---|
| Claude Opus 5.5 | 58 | $4 | $20 | Yes |
| Claude Sonnet 5.5 | 56 | $2 | $10 | Yes |
| Muse Wonder Full | 56 | — | — | No |
| Claude Fable 5.1 | 53 | $10 | $50 | Yes |
| GPT-6 Astra | 53 | $10 | $50 | Yes |
| Muse Wonder | 52 | $4 | $20 | Yes |
Wang said pricing would be "meaningfully below" OpenAI and Anthropic. That holds against Fable 5.1 and GPT-6 Astra at $10 input, but Opus 5.5 matches Muse Wonder on input, cached input and output — and outscores it. For the broader price ladder, see our pieces on Claude Haiku 5.5 at $0.10 per million tokens and GPT-6.1 Sol at $2.
On Meta's own 14-row comparison, Wonder Full takes seven rows and Wonder one (MemoryBench, 85.6 vs. 85.5). Wonder Full's biggest margins are on long autonomous coding (SWE-Marathon 57.5 vs. 46.3 for Opus 5.5) and legal agents (Harvey LAB-AA 13.1 vs. 4.2). Opus 5.5 wins ProgramBench 70.5 to 59.0. Two of Wonder Full's wins are on benchmarks Meta built, and Meta's model card says some rival scores were run in-house on a "best-effort" basis.
3. What Meta is testing with this model
| Question | What is known so far |
|---|---|
| Will people pay for Meta AI? | Muse Code at $50 and $200 a month; Muse app subscriptions in limited testing |
| Can Meta win enterprise API spend? | Only Wonder is on the API, also listed on OpenRouter |
| Safety | Meta claims its "strongest safeguards to date" against cyber misuse. Its safety card says it cannot rule out Wonder Full reaching its high-risk threshold for chemical and biological misuse without mitigations. Full safety report promised "soon" |
| Openness | Unlike the Llama line, weights are closed |
CNBC framed the release as a test of users' willingness to pay. Meta has funded its AI build-out from advertising; whether subscriptions and API revenue follow is the first read on returns from its data-center spending. In Meta's own cyber evaluation (ExploitGym), the model used the intended vulnerability on only 21 of 92 flags — meaning it often found another way around the test environment, the same family of behavior Anthropic disclosed about Claude agents bypassing tool limits.
4. What remains unconfirmed
- Independent testing: Only Artificial Analysis has published an independent score. Wonder Full, with no API, is harder for outsiders to test.
- Subscription prices: Reported Muse app prices differ ($20/$100 vs. $50/$200).
- Privacy complaints: Users alleged the Muse agent accessed private messages; Meta disputed this. Wonder's emphasis on long-term memory makes the question more pressing.
- Pre-release testing: Meta has not said whether the UK AI Security Institute tested the models before launch — a live question after the White House lifted its 17-day freeze on UK access on October 11.