Meta Muse Wonder: Meta's First Frontier AI Model Arrives at $4/$20
Fourteen months after restructuring around the Superintelligence Lab, Meta has entered the frontier model race. Muse Wonder and its stronger sibling Muse Wonder Full land at $4/$20 via the Meta Model API, with Active Memory that carries context across conversations and Active Focus that persists through errors — built specifically for agentic tasks that span days, not seconds.
- Frontier debut — Meta released Muse Wonder and Muse Wonder Full on October 12, 2026 — its first state-of-the-art models since forming the Superintelligence Lab under Alexandr Wang in August 2025.
- Agentic pricing — Muse Wonder costs $4 per million input tokens and $20 per million output tokens through the Meta Model API, roughly equivalent to GPT-6 Sol and Claude Opus 5.5. Muse Wonder Full is exclusive to Muse Code subscribers.
- Built for agents — The models introduce Active Memory (cross-conversation recall) and Active Focus (persistence through errors with clarification asks) — capabilities Meta says are essential for agents that navigate insurance claims, leases, and multi-step coding over long horizons.
On October 12, 2026, Meta's Superintelligence Lab announced Muse Wonder and Muse Wonder Full — the company's first genuine attempt at frontier-class AI models. Fourteen months after Scale AI founder Alexandr Wang was brought in to restructure Meta's AI efforts, the lab has moved from the lightweight Muse Spark to models that Meta says match Fable 5.1 and GPT-6 Astra on agent and coding benchmarks, while introducing two capabilities explicitly designed for long-horizon agentic work.
This article is based on Meta's official announcement (October 12, 2026), TechCrunch's launch-day reporting, and the published model card. Independent benchmark reproduction from Artificial Analysis is cited where available. Muse Wonder Full is not served via the public API — access is gated to Muse Code subscribers. All figures below are vendor-reported or from the cited independent evaluation; no hands-on testing was performed for this article.
What Meta actually shipped
Two models, two access paths:
- Muse Wonder — available today through the Meta Model API and Muse Code. Model ID
muse-wonder-1.0. Compatible with the OpenAI SDK (change one base URL). Cached input: $0.20 per million tokens. Context window: 1 million tokens. - Muse Wonder Full — stronger variant, Muse Code subscribers only, not served through the public API.
Both models will "soon power more of Muse" — Meta's personal AI agent that launched in September — rolling out first to subscribers. The company describes this as "a major step toward personal superintelligence for everyone."
The benchmark claims, and how to read them
Meta's model card reports state-of-the-art results on agent and coding evaluations:
- SWE-Marathon: 57.5% (Muse Wonder Full) vs 46.3% for Claude Opus 5.5 — a long-horizon software engineering benchmark.
- Artificial Analysis Intelligence Index v4.3.2: 56 (Muse Wonder Full) — behind Opus 5.5 at 58, ahead of GPT-6 Astra at 53. This is the independent evaluation to weight most heavily.
- MemoryBench: higher than Fable, Opus, and Astra — Meta's chosen benchmark for continual learning and memory construction across conversations.
Read the independent number first. Artificial Analysis scores Muse Wonder Full at 56 on its Intelligence Index — a respectable frontier showing, but not the "clear #1" the launch framing might suggest. The SWE-Marathon lead over Opus 5.5 is the strongest proprietary claim; it has not been independently replicated. The MemoryBench advantage is real per Meta's testing but the benchmark itself is less established than SWE-bench or Terminal-Bench.
Active Memory and Active Focus: the agentic differentiators
Meta's launch emphasizes two capabilities that are not standard benchmark rows:
- Active Memory — the model "brings information you need from across conversations, so you don't have to start from scratch each time." In practice this means cross-session context synthesis: preferences, project details, and facts persist and resurface when relevant.
- Active Focus — the model "continues working on a task even if it encounters errors or roadblocks... asks for clarification or permission to perform tasks." This is error recovery and initiative combined: the agent doesn't just fail and stop; it retries, pivots, and checks in before consequential actions.
These are the capabilities that matter if Muse becomes the brain of Meta's personal agent — an agent that books travel, navigates insurance claims, and builds apps over days of interaction. Standard chat benchmarks don't measure this. The real test is whether the agent actually completes multi-day workflows without human babysitting.
Infrastructure and training
The models were trained at Meta's Prometheus data center in central Ohio — a facility NBC4 called "the world's highest-capacity data center" when it came online earlier in 2026. Meta has not published parameter counts, architecture details, or training compute. The model card commits to a full safety evaluation report "soon," aligned with Meta's Superintelligence Scaling Framework.
Where this fits in the frontier landscape
October 2026 now has three frontier labs shipping models in the same price band:
| Model | Input /1M | Output /1M | Access | Agentic hook |
|---|---|---|---|---|
| Muse Wonder | $4 | $20 | API + Muse Code | Active Memory, Active Focus |
| GPT-6 Sol | $3 (promo) | $15 (promo) | OpenAI API | Dots, computer use |
| Claude Opus 5.5 | $4 | $20 | Anthropic API | Computer use, cache economics |
Pricing as announced October 2026. Sol's promotional rate guaranteed through at least November 21, 2026. Muse Wonder Full is API-exclusive to Muse Code; no public price published.
The structural story: Meta is no longer sitting out the frontier race. Whether Muse Wonder becomes the default brain for Meta's 3-billion-user surface (WhatsApp, Instagram, Facebook, Messenger) is the question that matters more than any benchmark. For developers, the API is live today at a competitive price, with OpenAI SDK compatibility lowering the switching cost.
For Meta watchers: the real product is not the model but the agent it powers. If Muse (the agent) ships Active Memory and Active Focus to billions of users, the model becomes infrastructure. The API is the proof point; the agent is the business.
For the frontier: three labs now occupy the $4/$20 tier with agentic positioning. The differentiation has moved from "who scores higher" to "whose agent actually finishes the job."
Bottom line: Meta has a frontier model, it's priced to compete, and it's built for the agent loop — not the chat turn. The independent Artificial Analysis score of 56 puts it in the conversation. The MemoryBench and SWE-Marathon claims need replication. For now: new option, live API, OpenAI-compatible, and the first frontier model explicitly architected for multi-day agentic continuity.
- Meta — Introducing Muse Wonder and Muse Wonder Full (official announcement, October 12, 2026)
- TechCrunch — Meta releases its first frontier model, Muse Wonder (October 12, 2026)
- Meta Research — Muse Wonder model card (benchmarks, safety, MemoryBench)
- Artificial Analysis — Muse Wonder Full provider benchmarking
- Bloomberg — Meta Says New Muse Model Hits AI Frontier After Aggressive Push (October 12, 2026)
Prices and benchmark figures are standard published rates as of October 23, 2026 and can change without notice. Some outbound links may be affiliate links — see our disclosure. Found an error? Tell us and we'll correct it in public.