Meta's first frontier model, priced at the frontier

Meta on Monday released Muse Wonder and Muse Wonder Full, which it describes as its most capable AI models and "a major leap toward personal superintelligence", according to a post from its research division. Muse Wonder is available now to developers through the Meta Model API and Muse Code, the company's coding agent, while Muse Wonder Full is limited to Muse Code subscribers. Meta said Muse Wonder will "soon" power its Muse personal agent app, rolling out first to subscribers; CNBC reported that the app's paid plans cost $20 or $100 a month. TechCrunch called it the company's first frontier model since Alexandr Wang restructured Meta's AI efforts in August 2025.

The pitch is aimed at two rivals by name. Reuters said the release was meant to let Meta compete more directly with OpenAI and Anthropic, and Meta's own comparison table sets the new models against OpenAI's GPT-6 Astra and Anthropic's Claude Fable 5.1 and Claude Opus 5.5. The models were known internally as Watermelon, Wang told CNBC. "Fourteen months ago, we began a ground-up overhaul of our artificial intelligence efforts," Meta wrote, adding that the models were trained at its "gigawatt-scale" Prometheus data centre, which TechCrunch located in central Ohio.

Meta's model card gives Muse Wonder a context window of one million tokens; it accepts text, images, video and audio and replies in text. On the Muse Wonder product page, the API price is $4 per million input tokens, $0.20 per million cached input tokens and $20 per million output tokens. Muse Wonder Full has no API price at all: it is available only on Muse Code's Power Usage plan, at $50 a month, and its Maximum Usage plan, at $200 a month, which Meta says carries 80 times the usage of its Everyday Usage plan.

The performance claims are Meta's. Its table shows Muse Wonder Full scoring 59.1 on Vals AI's Finance Agent benchmark, against 58.9 for Claude Fable 5.1, 58.6 for Claude Opus 5.5 and 53.5 for GPT-6 Astra, which is close to a tie. On Harvey's legal-agent benchmark, as run by Artificial Analysis, Muse Wonder Full scores 13.1 and the three rival models between 4.2 and 8.6. On DeepSWE, a long-horizon coding test, it scores 79.0 against 74.2 for Opus 5.5, 74.1 for Astra and 67.4 for Fable 5.1. On Meta's own memory test, MemoryBench, the smaller Muse Wonder scores 85.6, the highest in the table.

The same table shows where the models trail. Claude Opus 5.5 leads on Terminal-Bench, a coding test run in a command-line environment (59.6 against 57.1 for Muse Wonder Full), and on ProgramBench, which asks a model to rebuild a whole program from scratch (70.5 against 59.0). By our count, a Meta model posts the top score on 7 of the 14 rows Meta published, Opus 5.5 on six and GPT-6 Astra on one. Meta's widest coding margin, 57.5 against 46.3 for Opus 5.5 on the long-horizon SWE-Marathon test, comes with caveats: Muse Wonder Full was reported at the "xhigh" reasoning setting rather than the maximum used elsewhere, and the model card says the rival scores came through OpenAI's Codex and Anthropic's Claude Code, so the row "is not a controlled common-harness comparison".

Meta discloses that its method mixes sources. Some rows come from third parties, among them Artificial Analysis, Vals AI and Datacurve; others from Meta's own runs or from figures the rivals reported. Where Meta ran competitors' models itself, the model card says those runs "remain best-effort: prompts, tools, and runtime may not be optimally tuned for proprietary models." Several of the tests, including MemoryBench and the Personal Superintelligence Index, a composite of seven evaluations, are Meta's own or include Meta's own, and cannot be checked from outside.

One independent measure is already public. Artificial Analysis, which runs its own suite of tests, scores Muse Wonder Full at 56 on its Intelligence Index, ahead of GPT-6 Astra and Claude Fable 5.1 at 53 and behind Claude Opus 5.5 at 58, all at their maximum reasoning settings; Muse Wonder scores 52. Those are the figures Meta's table cites. Komposite found no other independent evaluation published by Monday evening, and none of the coding or legal results above has been reproduced by a party Meta did not choose.

Wang's account of the price was that the API would be "meaningfully below" leading models from OpenAI and Anthropic, CNBC reported. Against list prices, that depends on the comparison. Muse Wonder's $4 and $20 are, by our arithmetic, 60 per cent below the $10 and $50 that OpenAI charges for GPT-6 Astra and that Anthropic charges for Claude Fable 5.1. They are identical to Anthropic's price for Claude Opus 5.5, down to the $0.20 cached-input rate, and double OpenAI's $2 and $10 for GPT-6 Sol. Token use matters too: Artificial Analysis put the cost of one of its index tasks at $4.52 for Muse Wonder, against $3.26 for GPT-6 Astra and $5.98 for Opus 5.5.

For enterprises, the data terms deserve a reading. Meta's documentation says Muse Code subscriptions "default to Contributor mode", in which "you get more usage from your plan in exchange for letting Meta use your content to train future Meta models". It can be switched off, after which future content is not used for training, and both new models work in either mode. Meta's Muse Code pricing lists cheaper contributor rates for its older Spark models, marked "Used to improve our products"; it lists no contributor rate for Muse Wonder on the API.

The security disclosures are partial. Meta says Muse Wonder has its "strongest safeguards to date" against chemical, biological, radiological, nuclear and cyber misuse, but its full safety report is promised only "soon". The model card says that without mitigations, "we do not rule out the model reaching this threshold in the Chemical & Biological domain", meaning Meta's high-risk level, and that classifiers were added in response. CNBC reported that some Muse users have complained the agent accessed personal messages without authorisation, which Meta has disputed. "I think there are certainly instances users have authorized various communications or whatnot earlier in a conversation, and then may have forgotten about that," Wang said.

Wang described the training goal in terms of persistence. "When you do real-world tasks, you're going to get roadblocks and you're going to get things that come up or various setbacks," he told CNBC. "We specifically tried to train the model to be able to try other approaches instead of stalling and then checking back in with the user at any time." Meta calls this Active Focus, and pairs it with Active Memory, the ability to carry context about a user across conversations.

Investors welcomed the release. Meta's shares rose about 4 per cent in afternoon trading, Reuters reported, and closed at $742.96, up 3.4 per cent on Friday's close by our arithmetic from Yahoo Finance data. The release puts Meta in the same market for developers and coding seats in which OpenAI and Anthropic sell. Whether it takes business from them cannot yet be measured: Meta has published no usage figures for the API or Muse Code, and the cheapest route to its strongest model is a $50 subscription rather than a metered key.

Komposite reported last week on Manus, the agent start-up whose sale to Meta was unwound on Beijing's orders, and last month on how Microsoft meters frontier models from OpenAI and Anthropic on top of a seat licence. As we noted in September, the cost of an agent task can vary widely between runs, which is why a per-token price is only the start of a comparison.