Table of Contents
Meta Muse Wonder arrived on October 12, 2026 as Meta’s first frontier model — and the more interesting story is not the benchmark page. It is the door. Meta shipped two variants with two different access paths: Muse Wonder, served through the Meta Model API at $4 per million input tokens and $20 per million output tokens, and Muse Wonder Full, the stronger variant, which is not in the API at all. The only way to reach Wonder Full is a Muse Code subscription at $50 or $200 per month. That structure — a frontier model you can rent only through the vendor’s own agent product — is a pricing pattern the rest of the industry will be watching closely, and it is why Meta Muse Wonder deserves a closer look than the launch headlines gave it.
Key Takeaway
- 🚪 Two models, two doors: Meta Muse Wonder is in the Meta Model API and on OpenRouter; Muse Wonder Full exists only inside Muse Code’s $50 (Power Usage) or $200 (Maximum Usage) subscriptions.
- 💰 The price of the strong model: Meta’s best model is technically a feature of its coding agent — you cannot buy Wonder Full by tokens, only by plan.
- 📊 Independent scores so far: Artificial Analysis measured Muse Wonder at 52 and Wonder Full at 56 on its Intelligence Index — credible competitive numbers, not a new frontier crown.
- 🧭 What it means for you: If you build against APIs, Muse Wonder is a $4/$20 agentic option with a million-token context; if you want the strongest variant, you are buying into Meta’s agent ecosystem, not just a model.
The launch closes a fourteen-month rebuild. Meta’s Superintelligence Lab, restructured under Alexandr Wang since mid-2025, had shipped the lightweight Muse Spark and the consumer Muse agent but had not put a frontier-class model on the board. Muse Wonder and Muse Wonder Full are that first attempt, trained at Meta’s gigawatt-scale Prometheus data center, and the company frames them as state-of-the-art for personal and professional agentic tasks. The claim set is familiar; the packaging is not.
What Meta Actually Shipped: Muse Wonder, Muse Wonder Full
The Muse Wonder series is two models with one architecture story. Both carry Active Memory — the capability to hold on to the context that matters about you from one conversation to the next and pull the right details back for the current task — and Active Focus, the sustained-effort capability that tries alternative approaches, works through blockers, and checks in before acting on anything important. Meta’s research post also claims frontier coding performance and domain strength in law, finance, and health, measured against suites such as Harvey’s legal agent benchmark and Vals Finance Agent V2.
The two members of the Meta Muse Wonder family differ in strength and, critically, in distribution:
| Dimension | Muse Wonder | Muse Wonder Full |
|---|---|---|
| Meta Model API | Yes — muse-wonder-1.0 | No |
| OpenRouter | Yes | No |
| Muse Code subscription | Yes | Yes — Power Usage ($50/mo) or Maximum Usage ($200/mo) |
| API pricing | $4 / $20 per 1M tokens (in / out); $0.20 per 1M cache read | Not token-priced |
| Artificial Analysis Intelligence Index | 52 | 56 |
| Context window (OpenRouter listing) | 1,048,576 tokens | — |
Meta’s own blog positions both models as leading the field for personal and professional agentic work, with Muse Wonder Full as the stronger variant. The model card and benchmark charts are Meta-reported; independent numbers remain limited to aggregate indexes. Treat vendor benchmark charts as marketing until third parties reproduce them.
The Real Story: Meta Priced Its Best Model Into an Agent
Every major lab ties its strongest models to subscriptions to some degree — frontier features reach API tiers last. But Muse Wonder Full goes further: it is not available through the public API at any price. TechCrunch’s launch report, the Muse Code product page, and OpenRouter’s listing all agree on the split: the API gets Muse Wonder; Wonder Full is a subscriber capability of Muse Code, Meta’s coding agent.
The strategic logic is legible. Meta’s agent ecosystem — the Muse consumer app, connectors to email, calendars, shopping, and its new enterprise push — needs a reason to exist beyond commodity model access. A model that only lives inside Meta’s agent makes the agent the product and the model the moat. It also lets Meta price on value rather than tokens: a $50 plan that carries the strongest model is a bundle sale, not a meter. In the Meta Muse Wonder architecture, the subscription is not a wrapper around the model — it is the only road to the model.
The counterweight is developer friction. Teams that standardize on API-based workflows — orchestration layers, evaluation harnesses, cost dashboards — can buy Muse Wonder by the token but cannot bring Wonder Full into their stack. If Wonder Full genuinely leads on agentic tasks, its absence from the API is a structural compromise: the best model is fenced inside the vendor’s own harness. That is the real Meta Muse Wonder story: a capability split designed to pull buyers into the agent product rather than the meter.
How Strong Is Meta Muse Wonder, Really?
On independent measurement, the honest answer is: competitive, not crowned. Artificial Analysis — the most widely cited independent index — scores Muse Wonder at 52 and Wonder Full at 56 on its Intelligence Index, against frontier peers in the same range. Meta’s own claims run hotter: the company says the family scores higher on MemoryBench than Fable, Opus, or Astra, and posts its Personal Superintelligence Index numbers (54.3 for Wonder, 59.1 for Wonder Full). Meta’s benchmarks are vendor-reported; the Artificial Analysis numbers are the independent check.
The pricing lands mid-frontier: $4/$20 per million tokens is the same list price as Claude Opus 5.5 and roughly the frontier band for GPT-6-class models. OpenRouter’s listing adds texture for builders: a 1,048,576-token context window, multimodal inputs including text, images, video, audio, and PDF, tool calling, structured outputs, and ~71 tokens per second throughput on the Meta provider at launch.
For agentic work specifically, Meta emphasizes two properties that matter to practitioners: Active Memory for continuity across sessions, and Active Focus for sustained task execution — plus safety training around private information and misuse refusal. Those are the claims to watch as independent evaluations mature; memory and long-horizon focus are exactly the properties where 2026 benchmarks have shown current systems struggling.
The Enterprise Question: What Do You Get at Work?
Most launches now answer the “can my team use it?” question, and this one does — with a twist. Through the Meta Model API, an enterprise gets Meta Muse Wonder: frontier-band agentic performance, a million-token context, tool calling, and structured outputs at the same $4/$20 list as its peers. What an enterprise does not get at launch is the stronger variant, and what is not yet proven is the memory governance story at team scale — per-user memory scoping, admin visibility, retention controls, and audit trails are the questions Google answered explicitly at Gemini at Work, and they are the checklist any IT buyer should demand before wiring an agent into company data.
The sensible enterprise motion is a parallel pilot: run your own 20-to-50-task evaluation with Muse Wonder alongside your incumbent model, keep the volume modest until independent benchmarks and the promised safety report mature, and treat Wonder Full as a per-seat decision inside Muse Code if your developers already live in coding agents. That keeps the door open without betting the workflow on day-one pricing and claims.
How to Trial Muse Wonder Without Getting Burned
A disciplined one-week pilot costs little and answers more than any launch post. Five steps:
- Fix the test set before you look at the scores. Pick 20 tasks from your real backlog — drafting, extraction, refactors, research memos — and write down what “done” means for each. A benchmark you design on day one is the only one that describes your work.
- Run the same tasks through your incumbent. Muse Wonder’s $4/$20 pricing makes side-by-side costs trivial to log: tokens in, tokens out, cache reads, wall-clock time. Compute the real blended cost per completed task, not per token.
- Test the memory deliberately. The headline capability is Active Memory across conversations, so test it across conversations: brief it on a project in session one, then in session two ask for something that depends on the brief. Watch what it holds, what it drops, and whether a stale fact overrides a correction — that is exactly the experience-following failure current research warns about.
- Probe the refusal and privacy boundaries. Hand it ambiguous requests that mix personal and professional context. A model trained to handle private information carefully should narrow scope rather than guess; how it behaves at the boundary tells you more than any benchmark.
- Decide the tier honestly. If the API tier finishes your tasks at acceptable cost, you are done. If only the stronger variant clears your bar, recognize that you are then shopping for Muse Code — a $50–200 monthly seat in Meta’s ecosystem — and weigh that bundle against Claude Code or Codex plans doing the same job.
Who Should Use What — and Who Should Wait
The decision splits cleanly:
- API builders and tinkerers: Muse Wonder at $4/$20 is a legitimate frontier-band option with a million-token window and OpenRouter distribution. If your stack is model-agnostic, it costs little to benchmark it alongside Opus 5.5 and GPT-6 Sol on your own tasks — and the cache-read price ($0.20 per million) rewards agentic workloads with heavy context reuse.
- Teams already living in coding agents: Wonder Full is only reachable inside Muse Code. If you pay $50–200 monthly for a coding agent anyway, compare Muse Code’s plan against Claude Code and Codex plans on your actual repos. If Muse Code is not your tool of choice, Wonder Full’s strength is irrelevant — you cannot buy it anywhere else.
- Cost-sensitive agentic workloads: the $4/$20 pricing with cheap cache reads makes Muse Wonder worth a serious look for high-volume agent pipelines. Wait for independent task-level evaluations before committing production traffic.
- Anyone needing portability guarantees: a model that exists only inside one vendor’s agent product is a platform bet, not a utility purchase. If exit-cost matters to you, the API tier — or a different vendor — is the safer layer.
The Broader Pattern: Frontier Models Are Becoming Products
Watch what the Meta Muse Wonder launch does to the industry’s pricing grammar. For two years, the model was the product, the API was the distribution, and subscriptions were convenience wrappers. The two-door structure inverts that: the agent product is now the primary distribution for the strongest capability, and the API carries the second-best tier. If Wonder Full’s numbers hold up, expect competitors to test the same bundling — and expect API-only buyers to push back on the missing top tier.
There is also a governance thread worth tracking. Meta says it trained the family to handle private information carefully and to refuse misuse, with a full safety report promised after launch. A model designed to act across your messages, calendar, and documents is a privacy surface as much as a capability — and the promised safety report is where that gets tested. Until it lands, the privacy claims remain vendor statements.
For builders in Southeast Asia and everywhere else, the practical read is unchanged: the frontier band now includes Meta’s first entry at standard frontier pricing, with the strongest variant fenced inside the vendor’s own agent. Benchmark it on your tasks, watch the independent indexes, and read the bundle economics before committing — because this launch is as much about distribution as capability.
Frequently Asked Questions About Meta Muse Wonder Pricing and Access
What is Meta Muse Wonder?
Muse Wonder is Meta’s first frontier-class AI model, launched October 12, 2026 by the Superintelligence Lab alongside a stronger variant, Muse Wonder Full. Both are built for personal and professional agentic tasks, with named capabilities in memory (Active Memory) and sustained task execution (Active Focus).
How much does Muse Wonder cost?
The API tier, Muse Wonder, is priced at $4 per million input tokens and $20 per million output tokens, with cache reads at $0.20 per million. Muse Wonder Full has no API price because it is not served through the API — it comes with Muse Code’s Power Usage ($50/month) or Maximum Usage ($200/month) plans.
Can I use Muse Wonder Full through an API?
No. As of launch, Muse Wonder Full is exclusive to Muse Code subscribers. The Meta Model API and OpenRouter serve Meta Muse Wonder, not Wonder Full.
How does Muse Wonder compare to other frontier models?
Independent measurement from Artificial Analysis scores Muse Wonder at 52 and Wonder Full at 56 on its Intelligence Index — competitive frontier numbers. Meta reports stronger results on its own benchmarks, including MemoryBench wins over Fable, Opus, and Astra; those are vendor-reported and should be verified independently.
What is Active Memory?
Active Memory is Meta’s term for carrying important context across conversations and retrieving task-relevant details when needed. It pairs with Active Focus, which keeps the model working through blockers on long tasks. Both are the headline agentic capabilities of the Wonder family.
What is the context window and throughput?
OpenRouter’s launch listing reports a 1,048,576-token context window, multimodal inputs (text, images, video, audio, PDF), tool calling, structured outputs, and roughly 71 tokens per second on the Meta provider.
Is Muse Wonder available on OpenRouter?
Yes — Meta Muse Wonder is listed on OpenRouter at the same $4/$20 per million token pricing. Muse Wonder Full is not on OpenRouter.
Should my team adopt it now?
Run it through your own evaluation set first. The API tier is easy to pilot at frontier pricing, and cache-heavy agentic workloads get a discount via $0.20 cache reads. For Wonder Full’s strongest-tier capabilities, you would need to adopt Muse Code — a bundle decision, not a model decision.
Sources
Launch facts and capability claims for the Meta Muse Wonder family: Meta’s Muse Wonder research post and Meta’s announcement (October 12, 2026). Launch analysis and pricing: TechCrunch, the Muse Code product page, and the OpenRouter listing. Independent scores: Artificial Analysis. Vendor capability claims are Meta statements; independent index numbers belong to Artificial Analysis. For related WorldNgayon coverage, see our AI agent memory explainer, the Gemini agent launch analysis, and our frontier-pricing split coverage of GPT-6 Sol and Luna.
