Meta Releases Muse Wonder on Its Model API at $4 Input and $20 Output per Million Tokens, While Muse Wonder Full Is Limited to Muse Code Subscribers

Overview

Meta introduced Muse Wonder and Muse Wonder Full on October 12, 2026. In its newsroom post, the company calls them “our most capable models to date” and says Muse Wonder is available through the Meta Model API and Muse Code, while Muse Wonder Full is available to Muse Code subscribers. TechCrunch describes the release as “the company’s first attempt at a frontier model.” For developers, the practical details sit in Meta’s API documentation: a per-token price list, a preview label and a split between what the API serves and what only a subscription tool does.

What We Know

Availability. Meta’s Model API documentation lists Muse Wonder (muse-wonder-1.0) as available on pay-as-you-go in Standard mode only. It lists Muse Wonder Full as “Not available” on the Model API, and says the variant is available only in Muse Code with a subscription, on the Power Usage and Maximum Usage plans; Muse Wonder is on every plan. TechCrunch likewise reports that Wonder Full “will only be available to Muse Code subscribers and will not be served through the API.” The same documentation page says “Muse Wonder is in preview,” although the newsroom post says “available today.”

Pricing. Meta’s pricing page, read on October 13, lists Standard-mode rates for muse-wonder-1.0 of $4.00 per million input tokens, $0.20 per million cached input tokens and $20.00 per million output tokens. Contributor mode, the discounted tier that lets Meta use customer content to train future models, is not available for Muse Wonder on the Model API. The page says there is no long-context premium, and lists web search grounding at $2.50 per 1,000 search queries on top of token costs. It sets muse-wonder-1.0 limits at 3,000 requests per minute and 4,000,000 tokens per minute. TechCrunch says the price makes Wonder “roughly equivalent in cost to GPT-6 Sol or Opus 5.5”; that comparison is TechCrunch’s, and this article did not check those models’ prices.

The same pricing page lists Muse Spark’s Standard-mode rates at $1.25 per million input tokens and $4.25 per million output tokens. Those rates make Wonder about 3.2 times as expensive on input and about 4.7 times on output, by this article’s arithmetic. Meta’s documentation says Wonder and Spark take the same requests on the same endpoints, so moving a workload between them is a model ID change, and recommends Wonder for “long-running, complex tasks where capability matters more than cost per token.”

Specifications. The model card, which gives October 12, 2026 as the release date, lists text, image, video and audio input, text output and a context length of “1M tokens at inference.” The API documentation gives the context window as 1,048,576 tokens, lists PDF among the input types, and says reasoning effort runs from “minimal” to “max.” The model card also lists limitations, including that “Visual and spatial reasoning, especially 3D tasks, remains a known limitation.”

Benchmarks, as Meta reports them. The newsroom post calls the models state-of-the-art on personal and professional agentic tasks. According to TechCrunch, Meta’s testing in the model card shows “comparable benchmark scores to Fable 5.1 and GPT-6 Astra.” The model card says Meta’s results use max reasoning effort and compare against GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5 at max effort. It also describes how scores were chosen: where no comparable third-party or provider-reported result exists, Meta uses “the highest of our internally validated run, self-reported score, or official leaderboard score.” It says internally generated competitor runs “remain best-effort,” and that its SWE-Marathon row “is not a controlled common-harness comparison.” TechCrunch reports that, per Meta’s testing, Wonder scores higher than Fable, Opus or Astra on MemoryBench, which the model card describes as an internal benchmark. The score tables on Meta’s research page are graphics, so no figures are reproduced here.

Safety. The model card’s risk assessment uses Muse Wonder Full. Meta writes that without mitigations it does not rule out the model reaching the framework’s high-risk threshold in the chemical and biological domain, and it rates cybersecurity and loss of control as “moderate or lower” before mitigations. With mitigations, it reports “moderate or lower” across all domains. On the ExploitGym cyber evaluation, the card says Muse Wonder Full captured 92 flags, of which 21 used the intended vulnerability and 71 used unintended ones. Meta says a full report with additional safety evaluations “will be available soon.”

What We Don’t Know

• How long the preview label on the API documentation will remain, and what it means for service terms; the pages read do not say.
• What Muse Code subscription plans cost. The pages read for this article name the plans but not their prices.
• How Wonder performs in independent testing. Every capability figure above comes from Meta, and the numbers sit in charts rather than text.
• When Muse Wonder will arrive in the Muse assistant; the newsroom post says “soon,” to subscribers first.

Analysis

The documentation draws a line that the announcement does not: the cheapest route to Meta’s strongest model is a Muse Code subscription, not the API. Developers who call the API get Muse Wonder at a price several times Muse Spark’s, a line whose version 1.2 powered the beta Muse Code launch in August, when those same $1.25 and $4.25 pay-as-you-go rates were reported. Meta’s own guidance is to keep Spark for cost-sensitive work and use Wonder where capability matters more, which reads as a two-tier lineup rather than a replacement.