Meta’s Muse Wonder Returns Meta To The Frontier, With One Model You Cannot Buy By The Token

Posted on by
Contents show
A felt textile illustration of a tall plum block behind a stitched fence with a smaller mustard block in front, showing that Muse Wonder Full sits inside a subscription while Muse Wonder is sold by the token

Meta released two models on 12 October and split them across two very different ways of paying. Muse Wonder is sold by the token through the Meta Model API. Muse Wonder Full, the stronger of the two, is not sold by the token at all. It comes with a Muse Code subscription.

That single commercial decision is the most important thing about this launch for anyone building with AI, more important than the benchmark table that arrived with it. Meta spent fourteen months rebuilding its AI efforts around Meta Superintelligence Labs, and the result is a pair of models that are proprietary, agent-focused and priced at the premium end of the market rather than the value end where Meta used to compete.

This is what Meta has published about Muse Wonder, what it costs, where the numbers stop, and the one default setting in the subscription route that deserves to be read twice before any code leaves your machine.

1. What Meta actually released

Muse Wonder and Muse Wonder Full are the first frontier models from Meta Superintelligence Labs, the organisation Meta built after reorganising its AI efforts fourteen months ago. Meta says both models were trained at its gigawatt-scale Prometheus data centre, combining large-scale pretraining with reinforcement learning, and it describes the pair as the next step on its scaling ladder toward what it calls personal superintelligence.

The design goal is not conversation. Meta built the series for agents that hold onto context and keep working when a task gets difficult, and it names two capabilities as central to that:

  • Active Memory, which carries important context across conversations and retrieves the details relevant to the current task.
  • Active Focus, which keeps the agent pursuing an objective through obstacles by trying alternative approaches, and which Meta says is also trained to ask for clarification and to check in before important actions.

Meta's developer documentation describes the resulting behaviours in more familiar terms: multi-step reasoning over long horizons, end-to-end agentic task completion, coding agents that work across large repositories, tool use with recovery from failed calls, and multimodal reasoning over screenshots, charts, documents, images and video.

The lineage matters too. Muse Wonder follows the Muse Spark family, which Meta positioned as its efficient line. Where Muse Spark competed on price, Muse Wonder competes on capability, and Meta is now charging accordingly.

2. What the model costs, and what Meta does not charge for

Meta's pricing for the model that is actually on sale is straightforward:

ItemMeta's listed rate
Input4.00 dollars per million tokens
Output20.00 dollars per million tokens
Cached input0.20 dollars per million tokens
Web search grounding2.50 dollars per 1,000 search queries
Long contextNo premium, the same rate applies whether the context window is nearly empty or nearly full

That last line is the most interesting number in the table, because long context premiums have become the standard way to charge for agent work. A model that holds a million tokens of context at the same rate per token removes a cost cliff that appears in a great many agent designs, and for workflows that resend a large prompt on every turn, the combination of no long-context premium and a heavily discounted cached input rate is a meaningful difference. We covered the mechanics of that in the guide to inference caching, and Meta's approach follows the same logic: pay full price once, then pay almost nothing when the same prefix comes back.

Web search is billed separately at 2.50 dollars per 1,000 queries, which is worth modelling before you switch a research agent onto Muse Wonder, because search-grounded agents can issue many queries per task.

3. The stronger model is not for sale by the token

Meta's developer blog states that Muse Wonder Full is available exclusively by subscription in Muse Code, on the Power Usage and Maximum Usage plans. Meta says the Maximum Usage plan carries eighty times the usage limit of the Everyday plan, with the same five-hour and weekly resets. CNBC reported the two plans at 50 dollars and 200 dollars per month, and reported separately that the consumer Muse app carries subscriptions at 20 dollars and 100 dollars per month.

The consequence for a business is a change in how cost behaves, not only in how much it costs. A metered API bill scales with what your agent does. A subscription caps what your agent can do in a window and charges the same either way. For an unattended agent that runs continuously, the subscription is predictable and the ceiling is real; the moment you need that capability inside your own product, the subscription route stops being an option at all, because there is no published pay-as-you-go rate for Muse Wonder Full.

That is the gap enterprise buyers should notice. The stronger model can be tested inside Meta's coding assistant and cannot yet be embedded in customer-facing software, internal automation platforms or anything else you run outside Meta's tools.

4. The default setting that decides what you are trading

Meta runs two data-use arrangements, and the difference between them is the price of admission.

Standard mode charges the standard prices and states that your content is not used to train Meta models. Every model on Meta's pricing page is available in Standard mode, and on the Meta Model API, Contributor mode is not available for Muse Wonder at all. If you call muse-wonder-1.0 through the API, you are in Standard mode.

Contributor mode offers discounted pricing or more usage in exchange for permission to use your content, which Meta defines as the prompts, code, files and context you send, plus the model responses you get back, to train future Meta models. On the Model API it applies to the Muse Spark variants through dedicated model identifiers such as muse-spark-1.3-contributor.

The detail that matters for anyone considering the subscription route is in Meta's own documentation: in Muse Code, subscriptions default to Contributor mode, and the setting can be turned on or off in Muse Code's settings.

Read that in combination with the pricing. Muse Wonder Full is only reachable through Muse Code, and the Muse Code subscription arrives with Contributor mode switched on by default. A developer who signs up for the 50 dollar plan to try the strongest model Meta has built is, on the default settings, contributing the prompts, the code, the files and the answers to the training pipeline for future Meta models. That is a legitimate trade and Meta states it plainly. It is also a trade that many teams will make by accident, because the default is not where most people look.

Two questions are worth settling in writing before that happens. The first is whether there is an account-level setting that keeps Contributor mode off for every project rather than for the current session. The second is what happens to material that was submitted while Contributor mode was on if the setting is later switched off, because Meta's documentation describes how to change the setting without stating whether the change reaches material already submitted. I have not found an answer to either in the public documentation.

5. Where the models stand on independent testing

Artificial Analysis, which tests models independently, rates Muse Wonder Full at 56 on its Intelligence Index version 4.3 and the standard model at 52. Those figures appear both on the Artificial Analysis model pages and in Meta's own launch comparison, which is a useful coincidence: when a vendor republishes a third-party score accurately, the rest of its table deserves a more careful look than it would otherwise get.

Meta's claims around that index are strong. The developer blog states that Muse Wonder Full tops coding benchmark leaderboards including DeepSWE and SWE-Marathon in independent third-party runs and outperforms every other model evaluated, and that it competes with the best on agentic evaluations. Meta's own comparison table places it against Claude Fable 5.1, GPT-6 Astra and Claude Opus 5.5 across legal, financial, health, business workflow and long-running coding tasks.

The honest reading of that table is narrower than the headline. Meta describes its comparison as a mix of internally conducted evaluations, scores supplied by vendors and third-party results, which means the rows are not all measured under identical conditions. The Artificial Analysis index is the cleaner signal precisely because it applies the same ten evaluations to every model, and it puts the series in the leading group without putting it clear of the field. For a team deciding where to route work, that is the relevant conclusion: Meta is now a serious option on capability, and the choice between the leaders is a routing decision rather than an obvious one. Our earlier work on routing between models in production agent stacks covers how that decision usually gets made in practice.

6. What is missing from the announcement

Three gaps are worth naming, because each one affects a decision.

No downloadable weights. Meta's announcement describes availability through the Meta Model API and Muse Code, and it does not offer either model as a download. Meta built its reputation in open models, and Muse Glimmer, its smaller model, was released under Apache 2.0. The Muse Wonder series is a hosted product only, so self-hosting, private infrastructure deployment and direct modification are not on the table.

No published safety report yet. Meta says Muse Wonder offers its strongest safeguards to date against chemical, biological, radiological and nuclear risks and cyber misuse, that it preserves frontier cyber capabilities for defensive work under its Meta Superintelligence Scaling Framework, and that the model is its most aligned to date. The same post says a full report with additional safety evaluations will be available soon. That report was not published with the models, so the safety claims currently rest on the announcement rather than on the evidence behind it.

No pay-as-you-go route to the strongest model. This is the gap most likely to change, and it is the one to watch if your plans depend on embedding Muse Wonder Full in a product.

What the switch actually involves

Meta has made the technical migration deliberately easy. The documentation notes that a project built around the OpenAI SDK is one base_url change away from calling Muse Wonder, with the model identifier muse-wonder-1.0, and that rate limits apply per team rather than per API key. Background responses are capped at 600 submissions per minute per team, and Meta publishes guidance on exponential backoff for rate limit responses.

The practical effect is that trying Muse Wonder costs an afternoon. A team can send a representative workload through it, compare the output against whatever it uses today, and look at the billed tokens rather than the benchmark chart. Given that the API-available model and the subscription-only model are different products with different prices and different data terms, that test is worth running on the exact task you care about rather than on a general impression. The same discipline applies to the pricing comparison we published for Gemini 4 Argon and to long-horizon coding work like the overnight agent run we documented, where small differences in tokens per completed task decide the bill.

Where the numbers stop

  • I have not tested either model. Everything above comes from Meta's announcement, Meta's developer documentation and pricing page, the Artificial Analysis model pages, and CNBC's reporting of the plan prices.
  • The 50 dollar and 200 dollar Muse Code prices come from CNBC rather than from a Meta page I could read. Meta's developer blog confirms the plan names, the eighty times usage multiple and the reset windows without publishing the amounts.
  • Meta has not published a pay-as-you-go rate for Muse Wonder Full, and the announcement gives no date for one.
  • The benchmarks mix sources. Meta's own table combines internally conducted evaluations with vendor-supplied scores and third-party results, and the two Artificial Analysis scores are the cleanest figures in the launch material because one organisation applied one methodology.
  • Claims about safeguards and alignment are Meta's own, and the supporting report is described as forthcoming. Treat them as claims under test rather than as verified properties.

The launch also arrived alongside a consumer argument that touches the same subject. CNBC reported that some users of the Meta Muse app complained that the agent accessed personal messages without authorisation, which Meta disputed, and that Alexandr Wang said the company is monitoring for agent incidents and educating users about what they have authorised earlier in a conversation. That is the same pattern the governance research keeps finding: agents that hold context and act on it will sometimes act on consent that a person has forgotten giving.

FAQ

What is Muse Wonder?

Muse Wonder is Meta's frontier model series from Meta Superintelligence Labs, released on 12 October 2026 in two versions. Muse Wonder is available through the Meta Model API and in Muse Code, and Muse Wonder Full is available to Muse Code subscribers on the Power Usage and Maximum Usage plans. Both are built for agentic work, with Active Memory for context across conversations and Active Focus for persistence through obstacles.

How much does Muse Wonder cost?

Through the Meta Model API, Muse Wonder costs 4 dollars per million input tokens and 20 dollars per million output tokens, with cached input at 0.20 dollars per million and web search grounding at 2.50 dollars per 1,000 queries. Meta does not charge a long-context premium. Muse Wonder Full has no published pay-as-you-go rate and comes with a Muse Code subscription instead, reported at 50 dollars and 200 dollars per month.

Does Meta train on my code if I use Muse Wonder?

It depends on the route. On the Meta Model API, Contributor mode is not available for Muse Wonder, so API customers are in Standard mode, where Meta states that content is not used to train its models. In Muse Code, subscriptions default to Contributor mode, which permits Meta to train future models on prompts, code, files and outputs. The setting can be switched off in Muse Code, and Meta's documentation does not state whether switching it off affects material already submitted.

The short version

Meta returned to the frontier on 12 October with Muse Wonder and Muse Wonder Full, two proprietary agent models trained at its Prometheus data centre. Muse Wonder sells through the Meta Model API at 4 dollars per million input tokens and 20 dollars per million output tokens, with cheap cached input and no long-context premium. Muse Wonder Full, which Artificial Analysis rates higher, is locked to Muse Code subscriptions, and Muse Code subscriptions default to Contributor mode, which trades your prompts, code and files for training rights. The models cannot be self-hosted, the promised safety report is not out yet, and the strongest version has no pay-as-you-go price. For a small team the sensible move is to test the API model against a real workload, and to check the Contributor toggle before pasting a client repository into a coding agent.

Sources
Subscription Form