Meta launches Muse Wonder AI models, says they match OpenAI and Anthropic

Meta released two new AI models, Muse Wonder and Muse Wonder Full, on October 12, 2026. It calls them its most capable models yet and says they can rival leading models from OpenAI and Anthropic, citing its own internal benchmark testing. An independent index from Artificial Analysis gives a more measured picture. It scores Wonder Full 56, behind only Claude Opus 5.5 at 58.

The models come from Meta’s Superintelligence Lab, which Meta says has now built its most powerful AI systems. Meta is pitching them less as chatbots and more as engines for AI agents, meaning software that carries out multi-step tasks from start to finish. Developers can call Muse Wonder through the Meta Model API now. Wonder Full is kept behind a Muse Code subscription.

What Muse Wonder and Muse Wonder Full are

Meta offers the two models at different capability tiers. Muse Wonder is aimed at developers and businesses that need a frontier model for everyday tasks and coding. Muse Wonder Full is the more capable version, built for demanding technical work.

The release follows Muse Spark in April and the Muse personal AI agent in September. Muse Spark was the first model from Meta Superintelligence Labs. At launch, Meta described it as a natively multimodal reasoning model with tool use. Meta says it began a ground-up overhaul of its AI effort 14 months ago, and that both Wonder models were trained at its gigawatt-scale Prometheus data centre. The internal codename was Watermelon. According to 4SAPI, Mark Zuckerberg posted a watermelon emoji on September 2. During a livestream the weekend before release, he also said that “there is also a big watermelon on the menu”.

The launch follows a year of heavy investment by Zuckerberg in Meta’s AI work, including recruiting researchers from rival companies and spending on computing. Geo News reports that Meta’s stock rose nearly 4% in afternoon trading after the announcement.

What is new: Active Memory, Active Focus and a 1M-token context

Meta highlights two features:

• Active Memory retains relevant details from past conversations and brings them up when needed. This way users do not have to re-explain project context in every new session, which matters for work that runs over days or weeks.
• Active Focus keeps the model working on a hard task. It tries other approaches, recovers from errors and checks in before important actions. Meta says the model will still ask for clarification or check with the user before major actions.

Beyond those, Meta says the models can plan a sequence of steps, use digital tools and check their own work. On coding, they are described as working across large codebases, editing several files at once and running tests to verify their output. On long projects, they can hand subtasks to secondary agents. They can also operate desktop interfaces and web browsers, often called computer use. Meta pitches domain strength in law, finance and health, with examples such as navigating an insurance claim, understanding a lease or making sense of finances.

Reported specifications include a 1 million-token context window. A token is a small chunk of text, so the context window is how much material the model can consider at once. Inputs can be text, images, video, audio and PDFs, and output is text.

	Muse Wonder	Muse Wonder Full
Positioning	Everyday tasks and coding	Demanding technical work
Context window	1 million tokens (reported)
Inputs / output	Text, images, video, audio, PDFs / text (reported)
Artificial Analysis Intelligence Index	52	56 (high reasoning settings)
Access	Meta Model API, Muse Code	Muse Code subscribers only
API price (Meta)	$4 / $20 per million input / output tokens	No public API

Sources: Free Press Journal and Arab AI reporting on Meta’s announcement.

UX News points out that these are Meta’s descriptions of intended behaviour, not independent findings about reliability. Arab AI adds that the features do not mean the models will complete every task without mistakes.

How the models score in Meta’s tests and on Artificial Analysis

Meta’s own claims

Meta’s own benchmark chart compares the models with Claude Opus 5.5, Claude Fable 5.1 and GPT-6 Astra, with mixed results. According to 4SAPI, Meta says Wonder matches or slightly beats those three models on programming, financial and legal benchmarks. The Verge, as republished by WISE AI, notes that it is unclear how the models compare with Google’s Gemini 4 Argon.

The clearest lead in the reported figures is on long-running coding. On SWE-Marathon, a test of sustained software-engineering work, Wonder Full scores 57.5. Opus 5.5 scores 46.3 and GPT-6 Astra 42.5. The picture reverses on other tests. Opus 5.5 and GPT-6 Astra both beat Wonder Full on Terminal-Bench 4.0, a command-line task benchmark, and Opus leads on health questions. Datagrom reports that Wonder Full beats rivals on legal and coding benchmarks but trails Claude Opus 5.5 on health and computer-use tests.

Independent scores

On Artificial Analysis’s Intelligence Index, a third-party aggregate of many evaluations, Muse Wonder scores 52. Muse Wonder Full reaches 56 at high reasoning settings. That ties Claude Sonnet 5.5 and beats GPT-6 Astra, Gemini 4 Argon and Claude Fable 5.1, which all score 53. Only Claude Opus 5.5 scores higher, at 58. The index was recently rebuilt, as covered in our report on the new version. Arab AI notes that Artificial Analysis’s testing also shows clear gains over the earlier Muse Spark 1.3.

Arab AI adds that the gap narrows on tests of general knowledge and command-line tasks. 4SAPI’s summary: Meta’s vendor benchmarks claim parity or better on selected tests, while the independent index still puts the incumbent slightly ahead. In its words, the accurate framing is “close behind,” not “ahead”.

Where developers can use Muse Wonder, and at what price

Muse Wonder is available now through the Meta Model API and Muse Code, Meta’s coding assistant. Free Press Journal reports it is also on OpenRouter, but that is not confirmed by Meta in the sources. Meta prices Muse Wonder at $4 per million input tokens and $20 per million output tokens. Free Press Journal notes that this matches Claude Opus 5.5. Arab AI cautions that total cost depends on prompt size, output length and how heavily an application uses caching.

Muse Wonder Full has no pay-as-you-go API. It is limited to Muse Code subscribers. Free Press Journal reports this means the $50-a-month Power plan and the $200-a-month Maximum plan.

For the Muse consumer app, Meta plans a phased rollout that starts with paid subscribers. Meta also says it will use Wonder to build specialised models for all Muse users later. It has not shared specifics or a timeline for either. Meta’s Muse agent is also the focus of a separate partnership, covered in our report on Nvidia’s planned $3B investment in Meta.

What it means for builders

For teams already building agents, Muse Wonder adds another frontier-class option priced in the same bracket as Anthropic’s top model. Its reported strength is long, sustained coding work rather than short command-line tasks. That makes it most worth testing for multi-file refactors and long agent runs. It looks less suited to health workloads or computer-use automation, where Opus 5.5 leads in reported results.

The Meta Model API itself is new. As of August 10, 2026, WaveSpeedAI described it as a public preview framed for US developers, with Muse Spark 1.1 as the first supported model. The sources do not say whether its status or regional limits have changed since. WaveSpeedAI also notes that Meta says developers can point existing OpenAI SDK code at the Meta Model API. That lowers switching costs, but response fields, errors, tool semantics and refusal behaviour can still differ. Its advice: pin the exact model ID, log raw responses and verify quotas, pricing and data-retention terms before customer traffic goes through the API.

Active Memory raises its own product questions. UX News suggests testing whether users can see which remembered details influenced an action. It also says that on long tasks, progress visibility and the timing of approval requests become important. The announcement does not show how consistently these interactions work. Arab AI’s advice is to test the models against your own workflows, because benchmarks do not always reflect real-world use.

Open questions: safety, open source and the consumer rollout

Safety. Meta says Wonder has its strongest safeguards yet against chemical, biological, radiological, nuclear and cyber misuse, while keeping defensive cyber capability, under its Meta Superintelligence Scaling Framework. Meta also says a fuller safety evaluation report will follow.

Open weights. When Meta launched Muse Spark in April, it said it intended to open-source future versions. None of the sources describe open weights for either Wonder model. 4SAPI reads the launch as a strategic reversal for the company behind the open Llama line. It says Meta now states it will no longer prioritize open-source release. That is 4SAPI’s account, and other sources do not confirm the statement.

Access. Wonder Full is available only through subscription, and the Muse app rollout has no timeline. Many builders will therefore only be able to evaluate the smaller model directly for now.

What to watch next

• Meta’s promised fuller safety evaluation report for the Wonder models.
• Whether Wonder Full gets a pay-as-you-go API, and whether the reported OpenRouter listing is confirmed.
• Timing of the Muse app rollout to subscribers and the specialised models Meta plans for all Muse users.
• Independent results beyond the Intelligence Index, especially on computer use and health, where reports say Opus 5.5 leads.
• Rival releases. Anthropic announced Claude Haiku 5.5 the week before, and Google recently announced a Gemini agent for workplace tasks.