Meta Releases Muse Wonder - API Rates of $4 Input and $20 Output Match Opus 5.5, While Muse Wonder Full Is Limited to Muse Code's Top Two Plans

On October 12, 2026, Meta announced a new model, Muse Wonder, together with a higher-end variant, Muse Wonder Full¹. Meta calls the pair its most capable models and claims they reach state-of-the-art results on agentic tasks for personal and work use. Muse Wonder could be used from day one on the Meta Model API, Meta’s developer platform, and in Muse Code, the company’s terminal-based coding agent. Muse Wonder Full is reserved for Muse Code subscribers.

On the API, Muse Wonder is priced per million tokens at $4.00 for input, $0.20 for cached input and $20.00 for output³. Against the standard rates for Muse Spark, the model the Meta Model API offered until now ($1.25 input, $4.25 output), that is 3.2 times higher on input and roughly 4.7 times higher on output.

Only Wonder Reaches the API

The two models are sold through different channels. The availability table in the developer documentation comes down to this²:

Model	Meta Model API (pay-as-you-go)	Muse Code
Muse Wonder (muse-wonder-1.0)	Available (Standard mode only)	Pay-as-you-go and every plan
Muse Wonder Full	Not available	Power Usage and Maximum Usage plans

The documentation states plainly that raising Muse Wonder’s reasoning effort to the top “max” setting does not switch it to Full. The only route to Full is a Muse Code subscription on one of the two highest plans. Muse Code has four plans, Everyday Usage, High Usage, Power Usage and Maximum Usage, and a subscription can be used only from the Muse Code CLI⁵. That page does not list monthly prices for the plans.

On specifications, Muse Wonder takes input in text, image, video, audio and PDF form, produces text, has a context window of 1,048,576 tokens, and offers reasoning effort from “minimal” to “max”². For now it is labeled a preview. It is served on three endpoints, Responses, Chat Completions and Messages. Because it accepts identical requests through identical endpoints to Muse Spark, the documentation says moving between the two only requires changing the model ID.

The Same Numbers as Claude Opus 5.5

Put next to other vendors’ price lists, Muse Wonder lines up with Anthropic’s Claude Opus 5.5. Anthropic’s pricing page lists Opus 5.5 at $4 for input, $0.20 for cache hits and $20 for output, all per million tokens⁶.

Per million tokens	Muse Wonder	Claude Opus 5.5	Muse Spark (Standard)
Input	$4.00	$4	$1.25
Cached input	$0.20	$0.20	$0.15
Output	$20.00	$20	$4.25

The billing details are not identical, though. Claude’s price table has separate rates for cache writes ($5 for five minutes, $8 for one hour), while Meta’s pricing page gives no price for cache writes. Meta does spell out two other terms: the rate stays the same whether the context is nearly empty or nearly full, so there is no surcharge for long context, and web search grounding is billed on top at $2.50 per 1,000 queries³. Rate limits are set per team, and muse-wonder-1.0 gets 3,000 requests and 4 million tokens per minute, the same allowance as muse-spark-1.3.

With per-token rates at the same level, the comparison when choosing a model is likely to turn on factors other than price.

Even Meta’s Own Table Has Different Leaders by Task

On the model page, Meta publishes a comparison of Muse Wonder Full, Muse Wonder, Claude Fable 5.1, GPT-6 Astra and Claude Opus 5.5 across 14 benchmarks⁴. All figures below are as published by Meta.

On memory, which Meta’s blog promotes under the name Active Memory, Muse Wonder scores 85.6 and Muse Wonder Full 85.5 on MemoryBench, above Fable 5.1 (79.0), GPT-6 Astra (81.4) and Opus 5.5 (81.5). On SWE-Marathon, a measure of long-running software work, Wonder Full’s 57.5 (at xhigh) and Wonder’s 53.1 are also higher than the other three models, which range from 42.5 to 46.3. According to the model page, Full ranks first on agent and coding tests, DeepSWE and SWE-Marathon among them.

The same table also has rows where Muse Wonder falls behind:

• Terminal-Bench 4.0-AA: Opus 5.5 at 59.6 and GPT-6 Astra at 59.1, against 52.0 for Wonder
• ProgramBench: Opus 5.5 at 70.5, against 59.0 for Wonder Full and 50.0 for Wonder
• Artificial Analysis Index: Opus 5.5 at 58, against 56 for Wonder Full and 52 for Wonder

Since the model available on the API is Wonder rather than Full, it is worth checking which column is being read. The table has no column for Google’s Gemini 4 Argon, and The Verge noted that how the models compare with Argon is unclear⁸.

The API Only Offers the No-Training Option

Terms for data also differ between the API and a subscription.

On the Meta Model API, Muse Spark has a “Contributor mode.” In return for letting Meta use the prompts, code and model responses for training future models, the per-token rates drop sharply, to $0.10 for input and $0.20 for output in Muse Spark’s case³. In September, the latest Muse Spark 1.3 was added to that cheaper tier. Muse Wonder, however, has no Contributor mode on the API. Calling Wonder through the API means Standard mode only, in which, according to the documentation, submitted content and responses are not used to train Meta’s models².

Muse Code subscriptions work the other way: Contributor mode is switched on by default⁵. In that mode the plan’s usage allowance grows, and in exchange the content is used for training. The documentation says it can be turned off in settings at any time. On plans that include them, Wonder and Wonder Full can be used in either Standard or Contributor mode².

The same Muse Wonder therefore starts with opposite data terms depending on whether it is called with an API key or used through a Muse Code subscription. A team evaluating it needs to settle which of the two routes it will use first.

From Spark to a Top-End Model in Three Months

In July, Meta launched the Meta Model API alongside Muse Spark 1.1, and in August it released the terminal coding agent Muse Code in beta together with Muse Spark 1.2. In September it began offering Muse, a personal AI agent, in the United States.

In its announcement, Meta looks back on starting a rebuild of its AI work from the ground up 14 months ago and says Muse Wonder was trained at its gigawatt-scale Prometheus data center¹. TechCrunch reported that this is Meta’s first frontier model to come out since Alexandr Wang reorganized how the company runs its AI work in August 2025⁷. The publication framed the difference as Muse Spark targeting a lightweight, low-latency model, while Wonder is a larger model described as more capable.

Meta’s guidance on choosing between them suggests Wonder for long, complex work in which getting the job done well outweighs the per-token cost (coding changes across many files, agent runs lasting hours, coordinating sub-agents), and Spark for large volumes of work where speed or cost is the priority (sub-agent tasks, classification and similar)².

Open Questions Before Adopting It

Several points remain unsettled for anyone evaluating Muse Wonder:

• Preview status: No date has been given for general availability
• Regional availability: The documentation says model availability differs by country². Whether it can be used from Japan cannot be confirmed from the documentation
• Safety report: Meta claims Muse Wonder has its strongest protections yet against CBRN and cyber misuse, but says a complete report with further safety evaluations is still to come¹
• Use in Muse: Muse Wonder is due to power the Muse agent soon, starting with subscribers, but according to The Verge, Meta has given neither details nor a schedule⁸

With rates now level with Opus 5.5, whether to try Muse Wonder is likely to come down to two questions: whether a team’s work matches the areas where Meta claims an edge, such as memory and long-running development, and whether it will use the model under the API’s data terms or a subscription’s.

Sources

1. Muse Wonder - Meta AI Research official blog (October 12, 2026)
2. Models - Meta Model API - Meta developer documentation
3. Pricing and rate limits - Meta Model API - Meta developer documentation (pricing and rate limits)
4. Muse Wonder - Model page on Meta’s developer site (benchmark table)
5. Subscriptions - Muse Code - Meta developer documentation (Muse Code plans)
6. Pricing - Claude Platform - Anthropic official pricing page
7. Meta releases its first frontier model, Muse Wonder - TechCrunch (October 12, 2026)
8. Meta has some Wonder Full new AI models - The Verge (October 12, 2026)