Disclosure: this article is based on Meta’s model and pricing pages as reported by several outlets, plus Artificial Analysis index scores. We have not run Muse Wonder or Wonder Full on our own long-term workloads. Figures are attributed in the text.
Meta released Muse Wonder and Wonder Full on October 12, 2026. They are the most capable models Meta has shipped, they are closed, and they mark the clearest sign yet that Meta now intends to compete with Anthropic and OpenAI on paid frontier access rather than only on open weights.
The honest summary: Muse Wonder is priced exactly like Claude Opus 5.5 at $4 and $20 per million tokens, while its independent index score sits closer to GPT-6.1 Sol, which costs half as much. Wonder Full scores higher, but you cannot buy it by the token at all. And Muse Code subscriptions start in a mode that trains on your code. None of that makes Meta’s models bad. It does make them a harder sell than the launch coverage suggested.
Table of Contents
- What Muse Wonder Is
- Muse Wonder Pricing and Access
- How Muse Wonder Scores
- Price for Score: The Uncomfortable Comparison
- Active Memory and Active Focus
- The Contributor Mode Catch
- Where Muse Wonder Fits in Meta’s Plan
- How to Evaluate Muse Wonder Fairly
- Questions to Ask Before You Sign
- Who Should Use What
- FAQ
- Verdict
- Sources and Further Reading

What Muse Wonder Is
Muse Wonder is Meta’s new frontier model family, built by Meta Superintelligence Labs after Alexandr Wang restructured the company’s AI efforts. There are two versions. Muse Wonder is the standard model, available by API. Wonder Full is the stronger version, available only inside Muse Code, Meta’s terminal-based coding agent.
Both have a one-million-token context window, and neither has open weights. That second point matters, because Meta spent years as the champion of open models with Llama. The Muse line keeps the flagship closed and releases separate open-weight models underneath it: Muse Glimmer in August and Muse Spark Open on October 15, three days after this launch. Coverage of Spark Open was explicit that it is less advanced than Muse Wonder.
Meta pitches the models at expert domains such as law and finance, and at long coding and agent tasks. It also names two behaviours as the point of the release, Active Memory and Active Focus, covered below.
Several details remain unpublished. Yotta Labs noted that Meta’s developer post does not list rate limits, maximum output length or a web-search price, and lists audio input as unconfirmed. If any of those matter to your project, treat them as open questions.
Muse Wonder Pricing and Access
| Model | Where you get it | How you pay |
|---|---|---|
| Muse Wonder | Meta Model API, Muse Code, OpenRouter | $4 input, $0.20 cached, $20 output per 1M tokens |
| Wonder Full | Muse Code only | Power plan at $50/month or Maximum plan at $200/month |
| Muse app (consumer) | Meta’s consumer product | Separate $20 and $100 monthly tiers |
The per-token rates for Muse Wonder match Claude Opus 5.5 almost exactly, down to the $0.20 cached input. That is a deliberate positioning choice. Meta is saying it belongs in the premium tier.
Wonder Full has no API price at any level. MindStudio and others confirmed it is only available on Muse Code’s $50 Power Usage or $200 Maximum Usage plans. The practical consequence, as Rohit AI pointed out, is that paying for Full in Muse Code does not let your own application use Full. An app that wants Meta’s models needs its own API budget for the standard Muse Wonder.
Muse Code itself exited beta on September 1 with a starting plan at $5 a month, which The New Stack reported typically provides 10 to 50 requests every five hours. That is well below the headline subscription price of rival coding agents, though a low entry price says little about the cost of getting useful work done.
How Muse Wonder Scores
On Artificial Analysis’s Intelligence Index v4.3, Muse Wonder scores 52 and Wonder Full scores 56. Meta’s previous model, Muse Spark 1.3, scored 48. So both are clear steps up for Meta.
On GDPx, a benchmark of realistic professional work, Full scores 9.0% and standard Wonder 6.8%. Low absolute numbers are normal on that test, which is deliberately hard; the relative gap between the two versions is the useful signal.
For context, here is where some rivals sit on the same index, from published snapshots:
| Model | AA Index | Input / output per 1M |
|---|---|---|
| Wonder Full | 56 | Subscription only |
| Claude Sonnet 5.5 (max effort) | 56 | $2 / $10 |
| Muse Wonder | 52 | $4 / $20 |
| GPT-6.1 Sol (max effort) | 52 | $2 / $10 |
| Muse Spark 1.3 | 48 | Earlier model |
Index scores compress a lot of behaviour into one number, and different versions of the index are not directly comparable. Still, the pattern is hard to miss.
Price for Score: The Uncomfortable Comparison
Put the two tables together and the problem is obvious. Muse Wonder scores the same as GPT-6.1 Sol on the index and costs twice as much per token. Wonder Full matches Claude Sonnet 5.5 at max effort, a model you can buy at $2 and $10 by the token from four clouds.
There are reasonable counterarguments. Cost per task depends on how many tokens a model uses, not just its rate, and Meta has not published token-efficiency figures that would let anyone compare properly. Index averages also hide strengths. If Muse Wonder is unusually good at the long legal or financial tasks Meta targets, the average would not show it.
But the burden of proof sits with Meta. When a new entrant charges premium rates, buyers reasonably expect premium results, and the independent scores so far show competitive rather than leading performance. For most API buyers, Muse Wonder is worth testing, not worth switching to blind.
Active Memory and Active Focus
Meta’s two headline features address real problems in agent work.
Active Memory is the ability to carry what matters across separate conversations and pull back relevant details when a later task needs them. The goal is fewer repeated explanations. Anyone who has re-explained a project’s conventions to a coding agent every morning will see the appeal.
Active Focus is persistence. Meta describes it as keeping the model working toward the goal when a task gets hard: trying another approach, working through a blocker, recovering from errors, asking for clarification when needed, and checking in before doing anything important. It is aimed at the failure that costs most in practice, an agent that gives up at the first obstacle or claims a task is finished when it is not.
Both are Meta’s own descriptions and neither has been tested independently yet. Shattered made a fair point: Active Focus is genuinely hard to measure from outside, because it shows up in long-horizon task completion rather than a single benchmark score. Persistence also cuts both ways. OpenAI’s system card for GPT-6.1 Sol flagged unwanted persistence as a regression, so a model that never gives up is not automatically a good thing. The useful version knows when to stop.
The Contributor Mode Catch
This is the detail I would want every team lead to read before signing up. Muse Code subscriptions start with Contributor mode enabled. Meta’s subscription documentation says that mode gives more usage in exchange for permission to train future models on submitted prompts, code, files, context and responses.
O-Mega’s pricing guide adds that Meta’s own terms say you must not send confidential code in Contributor mode. In other words, the default setting on the subscription that unlocks Wonder Full is one that Meta itself says is unsuitable for confidential work.
The pay-as-you-go API is different. Rohit AI reports that Muse Wonder through the API runs in Standard mode only, which Meta says does not use submitted content for training. So the safer route for business code is the API, which means the standard model, not Full.
If you do use a subscription, switch Contributor mode off before sending anything proprietary, and check the setting after every plan change. Independent measurement cited by O-Mega suggests switching it off reduces your usage allowance, which is the trade Meta is offering.
Where Muse Wonder Fits in Meta’s Plan
Meta’s model line now has three layers. At the bottom is Muse Glimmer, a 30B Apache 2.0 model for local agents. In the middle is Muse Spark Open, an 816B open-weight model under a custom licence. At the top is Muse Wonder, closed and paid.
That structure looks a lot like what other labs do: open models to build an ecosystem and goodwill, closed flagships to earn revenue. It is a long way from the Llama-era argument that the frontier itself should be open. Reporting on the Muse Spark launch earlier in the year noted Meta planned to keep certain capabilities proprietary, and Wonder is where that plan lands.
For buyers, the upside is competition. A third serious US frontier vendor at the premium tier puts pressure on Anthropic’s and OpenAI’s prices. The downside is that Meta is new to selling API access, and its documentation, rate limits and enterprise terms are less mature than those of rivals who have done this for years.
How to Evaluate Muse Wonder Fairly
Given the gap between price and published scores, a fair evaluation needs to measure what Meta claims is different, not just accuracy on short tasks.
Pick ten long, multi-step tasks from your real backlog, the kind where agents usually stall or drift. Run each through Muse Wonder and through the model you use today, with the same tools and the same budget cap. Record four things: did it finish, how many tokens did it use, did it ask for clarification at sensible moments, and did its final report match what it actually did.
That last check is the most important. Active Focus promises persistence and honest check-ins. The way to test both is to compare the agent’s summary against the actual diff or output, task by task. A model that keeps working but reports success on unfinished work is worse than one that stops and says it is stuck.
Use the API in Standard mode for the test, so your evaluation tasks are not used for training. And give it a few weeks. Meta’s API is new, and early rate limits or latency may not reflect where the service settles.
Questions to Ask Before You Sign Anything
Enterprise buyers evaluating a new frontier vendor should get written answers to a short list of questions. Meta may well have good answers to all of them. The point is to have them on paper before your team builds a dependency.
What are the rate limits, and how do they scale? Meta has not published them. A model that performs well in a pilot but caps you at a low request rate in production is not usable for customer-facing work.
What is the maximum output length? Also unpublished. Long reports, large code changes and document generation all depend on it.
Which data is retained, for how long, and where? Standard mode does not train on your content, according to Meta. Retention for abuse monitoring, logging and regional processing are separate questions, and regulated industries need precise answers.
What is the deprecation policy? Anthropic now commits to at least a year of availability for new models. Ask Meta for an equivalent commitment, because rebuilding pipelines around a retired model is expensive.
Is there a path to the Full model by API? If the stronger model is the one that justifies the switch, you need to know whether your application will ever be able to use it, or whether it will stay subscription-only.
What support and service levels come with the API? Uptime commitments, incident communication and a named contact matter more than benchmark points once something is in production.
None of these questions are unusual. They are what any procurement team asks of a new cloud supplier. The difference with a brand-new API is that the answers may not exist yet, and finding that out before a migration is far cheaper than finding out after.
It is also worth asking how quickly the models will change. Meta shipped several Muse Spark point releases this year before Wonder arrived. Fast iteration is good for capability and hard on teams that need stable behaviour. A clear versioning policy, with pinned model IDs that do not change underneath you, is a basic requirement for anything in production.
Who Should Use What
| If you are | Start with | Why | Link |
|---|---|---|---|
| Curious about Meta’s frontier model | Muse Wonder via API | Standard mode, no training on your data | See pricing |
| A solo developer wanting Meta’s best | Muse Code Power, Contributor off | Only route to Wonder Full | See plans |
| Optimising cost per task | GPT-6.1 Sol or Claude Sonnet 5.5 | Similar index scores at half the rate | Compare |
| Wanting Meta models you can self-host | Muse Spark Open or Muse Glimmer | Open weights | See Spark Open |
Related on this site: for more context, see Muse Spark Open, Muse Glimmer and GDPx Benchmark Explained.
Frequently Asked Questions About Muse Wonder
When was Muse Wonder released?
Meta released Muse Wonder and Wonder Full on October 12, 2026. Muse Spark Open, a less advanced open-weight model, followed on October 15.
How much does Muse Wonder cost?
$4 per million input tokens, $0.20 cached and $20 output through the Meta Model API. Wonder Full has no per-token price and is only available on Muse Code’s $50 and $200 monthly plans.
Is Muse Wonder open source?
No. Both Wonder models are closed. Meta’s open-weight options are Muse Spark Open and Muse Glimmer.
How does it compare with Claude and GPT?
On Artificial Analysis’s index, Wonder scores 52 and Full 56, similar to GPT-6.1 Sol and Claude Sonnet 5.5 at max effort. Those rivals cost about half as much per token.
Does Meta train on my code?
In Muse Code’s default Contributor mode, yes, and Meta’s terms say not to send confidential code there. The pay-as-you-go API runs in Standard mode, which Meta says does not train on submitted content.
What are Active Memory and Active Focus?
Meta’s names for carrying context across conversations and persisting through hard tasks. They are Meta’s descriptions and have not been independently tested yet.
Verdict
Test Muse Wonder through the API on your hardest long-running tasks, because that is where Meta claims it is different. If Active Focus genuinely finishes work that other models abandon, the premium price could pay for itself.
The caveat is that the independent scores do not yet justify the price. It costs the same as Opus 5.5 while scoring like a model at half the rate, Wonder Full is locked behind subscriptions, and the default subscription mode trains on your code. Meta’s new features are still unverified claims.
Next step: open an API key, run ten long real tasks in Standard mode against your current model with a fixed budget, and compare finished tasks per dollar rather than benchmark scores.
Sources and Further Reading
- TechJack Solutions: Wonder vs Wonder Full, for pricing and the Active Focus description.
- MindStudio: Wonder Full pricing, for subscription-only access.
- Rohit AI: launch access, for Contributor and Standard modes.
- The New Stack: Muse Code pricing, for the coding agent’s plans.
- O-Mega: Muse Code plans, for the confidential-code warning.
On this site: compare coding tools in AI coding assistants in 2026, weigh vendors in AI models compared, protect agents with prompt injection and agent security, and control spend with LLM API cost optimization.


[…] on this site: for more context, see Muse Wonder and Claude Opus […]
[…] on this site: for more context, see Muse Wonder, Mistral Large 4 and Nemotron 3 […]
[…] Related on this site: for more context, see Claude Opus 5.5, GPT-6.1 Sol and Muse Wonder. […]