Meta's first frontier model, Muse Wonder, lands near the top, but the strongest version stays off the API

Meta finally has a model it is willing to call frontier. On Monday, October 12, 2026, Meta Superintelligence Labs released Muse Wonder and Muse Wonder Full, which Meta's announcement describes as "our most capable models to date." TechCrunch called them Meta's first attempt at a frontier model, and they are the first release of that class since Alexandr Wang took over and restructured the company's AI work in August 2025.

The first independent numbers back up the claim that Meta is now in the frontier pack, but not that it leads it. There is also a catch in how Meta is shipping the two models: the stronger one, Wonder Full, is not available through the API at all.

What Meta actually shipped

There are two models. Per Meta's announcement, Muse Wonder is available now to developers through the Meta Model API and Muse Code, while Muse Wonder Full is available only to Muse Code subscribers. Both models will also power Meta's consumer Muse agent, which Meta says will get Wonder "soon," rolling out to subscribers first. Meta's announcement says the models build on Muse Spark, its earlier lightweight, low-latency model. Axios reported Muse Spark's debut in April as the first model released under Wang.

According to Meta's model card, dated October 12, both models take text, image, video and audio as input, produce text, and have a context window of 1 million tokens. Meta has not disclosed parameter counts. Both models are proprietary. VentureBeat notes that neither has downloadable weights, a notable shift for the company that built its AI reputation on the open-weight Llama family. TechCrunch, citing Meta's post, reports that the models were trained at Prometheus, Meta's data center in central Ohio, which came online earlier this year.

Meta pitches the models on two features. "Active Memory" is meant to carry information across conversations and bring back details relevant to the task at hand. "Active Focus" is meant to keep the model working toward a goal when a task gets hard: trying other approaches, recovering from errors, and asking for clarification or checking in before important actions. In a launch video quoted by TechCrunch, Meta research scientist Kavya Kopparapu said: "We put a tremendous amount of effort into making sure your model can remember what's important to you."

Pricing and access

VentureBeat's reporting has the API details. The model identifier is muse-wonder-1.0, and pricing is $4 per million input tokens, $20 per million output tokens and $0.20 per million cached input tokens. Built-in web search costs $2.50 per 1,000 queries. Rate limits are 3,000 requests and 4 million tokens per minute. The API supports interfaces compatible with OpenAI's and Anthropic's developer tooling, which lowers the cost of trying it in an existing stack.

That per-token price matches Claude Opus 5.5. According to Artificial Analysis's published price comparison, GPT-6 Astra and Claude Fable 5.1 cost $10/$50, while Claude Sonnet 5.5, Gemini 4 Argon and GPT-6.1 Sol cost $2/$10. Muse Wonder is a big step up from Meta's own Muse Spark 1.3, which costs $1.25/$4.25.

Wonder Full is sold differently. VentureBeat reports that Muse Code has a $50-a-month Power Usage tier and a $200-a-month Maximum Usage tier, which come with 20 and 80 times the usage allowance of the Everyday plan. Usage limits apply, and access is through the Muse Code assistant, not general API calls. If you want to build a product on Meta's best model, you can't do that today.

What the independent numbers say

Artificial Analysis published its evaluation the same day. On version 4.3 of its Intelligence Index, a composite of ten evaluations run at maximum reasoning effort, Muse Wonder scored 52 and Muse Wonder Full scored 56. Muse Spark 1.3 scored 48, six weeks earlier. For comparison, Artificial Analysis and VentureBeat put Claude Opus 5.5 at 58, Claude Sonnet 5.5 at 56, and GPT-6 Astra, Gemini 4 Argon and Claude Fable 5.1 at 53 each.

That puts Wonder Full near the top of the table, tied with Sonnet 5.5 and ahead of several flagship models, but behind Opus 5.5. The standard Wonder model, the one you can actually buy by the token, lands just below the 53-point cluster. Artificial Analysis says Wonder Full places in the top three on six of the ten evaluations in the index. It ranks first on CritPt (33%) and GDP.pdf (33%), but the firm calls both leads narrow.

On agentic coding, Artificial Analysis's Coding Agents Index gives Muse Wonder Full running inside Muse Code a score of 68, tied with Claude Code running Sonnet 5.5 and ahead of Claude Code with Opus 5.5 at 66. Plain Muse Wonder scores 63. Muse Code running Muse Spark 1.3 scored 54.

The weak spot is knowledge. On Artificial Analysis's AA-Omniscience knowledge evaluation, Wonder Full scores 30 and Wonder 25. Claude Opus 5.5 scores 46, GPT-6 Astra 43 and Gemini 4 Argon 42. Wonder Full's hallucination rate fell from 33% to 23%, which the firm says is the second-lowest among models scoring 45 or more on the index, behind Gemini 4 Argon at 15%.

Cost per task matters more than the sticker price, and here Muse Wonder isn't a bargain. Artificial Analysis estimates that running Wonder averages $4.52 per Intelligence Index task, 2.8 times Muse Spark 1.3. It is cheaper than Opus 5.5 at $5.98 and Fable 5.1 at $7.63, but more expensive than GPT-6 Astra at $3.26 and Gemini 4 Argon at $1.99. Most of the cost is input tokens. Wonder consumed about 7.69 million input tokens per task, while output stayed around 54,000 tokens.

Meta's own benchmark claims

Meta's model card makes more aggressive claims, and they need context. VentureBeat lists the card's head-to-head results for Wonder Full against Opus 5.5 and GPT-6 Astra. On SWE-Marathon, Meta reports 57.5 against 46.3 and 42.5. On DeepSWE 1.1, it reports 79.0 against 74.2 and 74.1. On the Harvey LAB-AA legal benchmark, it reports 13.1 against 4.2 and 8.6. The same card shows Wonder Full losing on Terminal-Bench 4.0-AA (57.1 against 59.6 and 59.1) and on HealthBench Pro (62.3 against 65.6 and 65.3), and essentially tied on JobBench.

VentureBeat points out that Meta's comparison mixes internal evaluations, vendor-supplied scores and third-party results, and that some coding comparisons used different tools. TechCrunch notes that Meta's MemoryBench claim, that Wonder beats Fable, Opus and Astra at memory construction, comes from Meta's own testing. Meta's card also lists its own limitations, according to third-party summaries: weak visual and spatial reasoning, especially in 3D, and mistakes on unfamiliar multi-step problems. The legal scores are also low in absolute terms. The best result on Harvey LAB-AA is 13.1.

The data-handling fine print

Developers should read the terms before pointing Muse Code at a private repository. VentureBeat reports that the pay-as-you-go API runs in a "Standard" mode where Meta says submitted content isn't used for training. Muse Code subscriptions default to "Contributor" mode, where Meta may use prompts, code, files and outputs to improve its models. Users can opt out, but the opt-out only covers future submissions, and Meta's terms prohibit submitting confidential, sensitive or personal information under Contributor services.

Zero data retention is available to approved API organizations. According to VentureBeat, it doesn't support server-side conversation state, background requests, file uploads or built-in web search. Those are the features Active Memory and the agent workflows depend on, so the most private setup is also the least capable one.

On safety, Meta's announcement says the models were trained to avoid sharing private information without permission and to refuse genuinely harmful requests. It says they were developed under Meta's Superintelligence Scaling Framework, and that a full report with additional safety evaluations is "available soon." That report had not been published at launch.

Why it matters

For the last year the frontier race has been a three-way contest between OpenAI, Anthropic and Google. Muse Wonder doesn't change the order at the top. Opus 5.5 still leads Artificial Analysis's index, and Meta's model is noticeably weaker on knowledge. But it does give Meta a credible fourth entry, a little over a year after the restructuring, built on its own training infrastructure. The bigger change is strategic. Meta is now shipping its best model as a closed product, held back from its own API and sold inside a subscription coding tool. That looks a lot more like Anthropic's playbook than Llama's.