Meta has released its first frontier artificial intelligence model. Muse Wonder — and a larger sibling, Muse Wonder Full — arrived on 12 October 2026 from the company's Superintelligence Lab, and Meta claims both are built for "personal and professional agents" and perform level with the top models from OpenAI and Anthropic on agentic tasks. It is the most consequential AI announcement Meta has made since it tore up its AI strategy fourteen months ago — and the first real test of whether the company can compete at the frontier rather than merely near it.
What Meta announced
In a newsroom post on Monday, Meta introduced the Muse Wonder series as its most capable models to date, built on the groundwork of the lightweight Muse Spark model that arrived in September. The company says the new models deliver state-of-the-art performance on agentic work: long-running tasks where an AI uses tools, works across several applications, and keeps going until the job is done. Examples in Meta's announcement ranged from navigating an insurance claim and understanding a lease to frontier coding.
The release is also a business move. Meta wants the new models to power Muse, the personal agent it launched in September — a product whose global download growth has reportedly slowed, according to the Wall Street Journal. Rolling out first to subscribers, Muse Wonder is meant to make that agent smart enough to keep users coming back. The models were trained at Meta's gigawatt-scale Prometheus data centre in Ohio, and neither is released with publicly downloadable weights — a decisive break from the open-weight Llama era.
Two models, two ways to get them
Muse Wonder (model identifier muse-wonder-1.0) is the version developers can actually use. It is available pay-as-you-go through the Meta Model API — also listed on OpenRouter — and inside Muse Code, Meta's coding tool, on every plan. API pricing is set at US$4 per million input tokens and US$20 per million output tokens (use our currency converter to see what that costs in rupees), matching the list price of Anthropic's Claude Opus 5.5. Rate limits run at 3,000 requests and 4 million tokens per minute per team.
Muse Wonder Full is the bigger, more capable sibling — and most of the headline benchmark numbers come from it. But it is not available through the API at all. The only way to use it today is inside Muse Code on the Power Usage or Maximum Usage subscription plans. For developers weighing a switch, that matters: the standard model's published numbers are one thing, but the strongest results belong to a version sold only by subscription. The model offers a context window of about one million tokens, accepts text, images, video, audio, and PDF inputs, and supports adjustable reasoning effort from minimal to maximum, alongside tool calling, structured output, search grounding, and computer use.
Active Memory and Active Focus, explained
Meta built both models around what it says an agent needs to do long tasks for a person — and gave the two headline capabilities names. Active Memory is the ability to carry important context about a user across conversations and pull up the details that matter for the task at hand, so a user does not have to repeat information when working on ongoing projects. Active Focus is persistence: keeping going toward a hard goal, trying other approaches when blocked, recovering from errors, asking for clarification when needed, and checking in before taking important actions.
The framing targets two of the most common failure modes of AI agents today: losing context over long sessions, and drifting away from the original objective. Meta also points to tasks in law, finance, and health as part of its evaluation — understanding a lease, making sense of finances — and describes frontier coding performance tested on long-running software work with a benchmark it calls SWE-Marathon. Memory, notably, is becoming a theme across the agent tooling world: the industry is converging on the idea that agents need something beyond a context window to be genuinely useful.
What the benchmarks actually show
Meta's own charts show its two models on top in roughly half of the tests it chose to publish, with Anthropic's Claude Opus 5.5 leading almost all of the rest. The standout internal number is 57.5 per cent for Muse Wonder Full on SWE-Marathon, against 46.3 per cent for Claude Opus 5.5 — though that is Meta's benchmark, run by Meta, and should be read with that in mind.
Independent testing tells a closer story. The Artificial Analysis Intelligence Index scores Muse Wonder Full at 56, behind Claude Opus 5.5 at 58 and ahead of GPT-6 Astra at 53 — competitive, but not a decisive lead. For a company trying to prove it belongs at the frontier after a long absence, "level with the leaders" is arguably the point: Meta needed to show it could play the game, not necessarily that it had already won it. And Meta has not yet published the safety evaluation report it promised, so its safety claims — including that Muse Wonder is its "most aligned model to date" — cannot yet be checked against full test results.
Why Meta needed this
The release comes fourteen months after Meta began rebuilding its AI operation, a period that started with the creation of the Superintelligence Lab and a shift away from the open-weights strategy that defined the Llama years. The earlier Muse Spark was deliberately a lightweight, low-latency model; Muse Wonder is the first output of the rebuilt effort aimed squarely at the frontier.
