Meta releases its first frontier model, Muse Wonder

Meta has finally joined the frontier AI race in earnest. On Monday, 12 October 2026, the company's Superintelligence Lab unveiled Muse Wonder and Muse Wonder Full, a pair of large language models that Meta describes as its first genuine attempt at frontier-class AI. The release marks a decisive turn for a lab that has spent more than a year reorganising under former Scale AI chief Alexandr Wang while rivals OpenAI and Anthropic shipped one state-of-the-art system after another.

Where Meta's earlier Muse Spark was a lightweight model tuned for low latency, the Wonder family goes the other way: bigger, more expensive to run, and aimed squarely at complex work such as coding, long-horizon agent tasks and professional domains like law, finance and health. Meta's own testing puts both models roughly on par with Fable 5.1 and GPT-6 Astra — the current leaders' flagships — and the company is pitching a distinctive strength in memory, which it calls "active memory". The move also confirms a strategic shift: unlike the Llama generation, Muse Wonder is a closed model, available only through Meta's own channels.

What Meta actually launched

Muse Wonder is the standard model. It is available through the Meta Model API (also on OpenRouter as meta/muse-wonder) and inside Muse Code, with a one-million-token context window, multimodal input covering text, images, video, audio and PDFs, and adjustable reasoning effort from minimal to maximum. It supports tool calling, structured output, search grounding and computer-use style operation — in other words, everything needed to drive an autonomous agent.

Muse Wonder Full is the larger sibling, and it is deliberately harder to get. Meta is not serving it through the API at all; access is restricted to Muse Code subscribers on the Power (US$50 per month) and Maximum (US$200 per month) plans. Most of the headline benchmark numbers Meta released were produced by the Full model, which makes the restriction a familiar playbook from the industry: keep the most capable version close, and let it anchor a paid developer product.

Both models will also power the Muse agent, Meta's long-running personal assistant effort. That connection is not incidental. Meta has spent years arguing that its endgame is "personal superintelligence" — an assistant that works for you over days, weeks and months — and the Wonder launch post explicitly frames the pair as "the next step on our scaling ladder and a major leap toward personal superintelligence."

Benchmarks: where Wonder stands

Meta's model card claims Wonder and Wonder Full achieve benchmark scores comparable to Fable 5.1 and GPT-6 Astra. Third-party numbers paint a picture of a genuine frontier contender, though not a clear winner:

• Artificial Analysis Intelligence Index: Muse Wonder Full scores 56 and standard Muse Wonder scores 52, against 58 for Claude Opus 5.5 and 53 for GPT-6 Astra. Spark 1.3, released roughly six weeks earlier, sat at 48.
• Coding: Wonder Full reaches 68 on the Coding Agents Index, matching top Anthropic models, and scores 79.0 on DeepSWE 1.1. On Terminal-Bench 4.0 it leapt from 33% to 57% success.
• Agentic tasks: Meta's own SWE-Marathon figures put Wonder Full at 57.5 against 46.3 for Claude Opus 5.5 — though that is the company's own test, and independent reads are closer.
• Weak spots: factual knowledge lags rivals, with just 30 points on the AA-Omniscience test, suggesting Wonder is stronger at reasoning and doing than at recalling facts.

The honest summary is that Meta has built a frontier model that is competitive with the leaders on the tasks that matter most for agents and coding, without clearly beating anyone. That is still a remarkable turnaround for a lab that, a year ago, was widely viewed as having fallen behind.

The memory pitch: active memory and active focus

The most distinctive part of Meta's pitch is memory. The company has placed heavy emphasis on the models' ability to retain and synthesise information supplied by the user over long periods — a property Meta brands "active memory". Research scientist Kavya Kopparapu said in the launch materials that the team had invested heavily in making the model remember what matters to the user across days, weeks, months and beyond.

To back the claim, Meta highlighted MemoryBench, a benchmark that stresses continual learning and memory construction. According to Meta's testing, Muse Wonder now outscores Fable, Opus and Astra there — the one area where the company claims an outright lead. The bigger model adds a second feature, "Active Focus", designed to keep agents working through difficult goals, trying different approaches and checking in before taking consequential actions rather than drifting off-task.

This is a deliberate differentiator. The frontier race has largely converged on raw reasoning power; everyone is good at maths and coding now. Meta is betting that the next battleground is persistence — an assistant that remembers you and stays on task. For agent workloads, where models routinely lose track of instructions across long tool-use chains, that is a bet with real logic behind it, even if MemoryBench is a company-chosen metric that deserves independent scrutiny.

Pricing and the end of the open-weights era

Frontier capability comes with frontier pricing. Accessed through the Meta Model API, Wonder costs US$4 per million input tokens and US$20 per million output tokens, roughly in line with GPT-6 Sol or Claude Opus 5.5, and well below the US$10/US$50 listed for GPT-6 Astra. In real-world cost terms, Artificial Analysis estimates US$4.52 per benchmark task for Wonder — nearly triple the US$1.60 for the previous Muse Spark 1.3, cheaper than Anthropic's lineup and pricier than GPT-6 Astra or Gemini 4 Argon.

The pricing tells the bigger story. Meta spent years positioning itself as the open-weights champion while Google, OpenAI and Anthropic charged for frontier access. Wonder's price list, combined with closed weights and a flagship model reserved for paid subscribers, shows that era is over. The Llama generation made Meta the industry's open-source standard-bearer; the Muse generation makes it just another frontier lab with a meter running.

That has a commercial logic. Training at the scale of Meta's gigawatt-scale Prometheus data centre in central Ohio — where the Wonder models were reportedly trained — costs enormous sums, and Meta's investors have been pressing for the AI buildout to show returns. A premium-priced API and a subscriber-anchored flagship are how you monetise a frontier lab.

Why this launch matters for the AI race — and for markets

Muse Wonder lands about fourteen months after Alexandr Wang began restructuring Meta's AI efforts, a period in which the company reorganised, released the lightweight Spark and the Muse agent, but conspicuously avoided a flagship model release. The launch is therefore as much a credibility event as a product event: Meta is signalling to developers, investors and rivals that the catch-up phase is finished. Mark Zuckerberg announced the models in a thread on X declaring that Wonder "brings us to the frontier" — a claim the independent benchmark numbers make plausible, if not yet triumphant.

For developers, the practical questions are cost, reliability and data terms. Wonder is priced to compete with the leaders rather than undercut them, so teams will switch only where it demonstrably wins on their own tasks. For the broader market, every frontier launch reshuffles expectations about who captures AI value. Mega-cap tech stocks have been known to swing sharply on model announcements, and investors watching that volatility often keep an eye on diversifiers — from today's gold rates to currency moves — as a counterweight to tech exposure.

The deeper significance is structural. With Meta now fielding a competitive closed frontier model, the era in which only two or three labs defined the state of the art is giving way to a genuine four- or five-way contest. More competition at the frontier should mean faster capability gains and — eventually — better prices. Whether Meta's memory-first differentiation proves to be a real moat or a marketing label will be settled not by model cards, but by how Wonder performs in developers' hands over the coming months.

FAQs

What is Muse Wonder?

Muse Wonder is a family of large language models released by Meta's Superintelligence Lab on 12 October 2026. It comes in two versions: the standard Muse Wonder, available through the Meta Model API and Muse Code, and the larger Muse Wonder Full, restricted to Muse Code subscribers. Meta presents both as its first frontier-class models — systems competitive with the best offerings from Anthropic and OpenAI — with a particular emphasis on long-term memory and agent-style task execution.

How does Muse Wonder compare to Claude Opus 5.5 and GPT-6 Astra?

On the independent Artificial Analysis Intelligence Index, Muse Wonder Full scores 56, standard Muse Wonder 52, Claude Opus 5.5 58 and GPT-6 Astra 53 — so Wonder is close to the leaders without clearly beating them. Meta's own testing claims comparability with Fable 5.1 and GPT-6 Astra, and shows Wonder Full ahead on some coding and agent benchmarks such as SWE-Marathon. The balanced reading is that Wonder is a credible frontier model that trails the very best by a small margin on general intelligence, while leading on Meta's chosen memory benchmark.

How much does Muse Wonder cost to use?

Through the Meta Model API, Muse Wonder costs US$4 per million input tokens and US$20 per million output tokens, with a one-million-token context window. That puts it roughly level with GPT-6 Sol and Claude Opus 5.5, and above GPT-6 Astra's listed pricing. Independent estimates put the real-world cost at about US$4.52 per benchmark task, nearly triple the cost of the previous Muse Spark 1.3. Muse Wonder Full has no public API pricing because it is not served through the API at all.

Is Muse Wonder open source?

No. Unlike Meta's earlier Llama models, whose weights were released for anyone to download and run, Muse Wonder is a closed model available only through Meta's API, Muse Code and Meta's apps. The company has not said whether it will ever release the weights. This represents a strategic pivot: Meta spent years as the industry's open-weights champion, and the Wonder launch — with closed weights and premium pricing — signals that the company now intends to monetise its frontier models directly.

What is "active memory" and why is Meta emphasising it?

Active memory is Meta's term for the Wonder models' ability to retain and synthesise information a user provides across conversations stretching over days, weeks and months — essentially, an assistant that remembers you rather than starting each session from scratch. Meta claims Muse Wonder outscores rival models on MemoryBench, a benchmark for continual learning and memory construction. The emphasis matters because frontier models have largely converged on raw reasoning ability; Meta is betting that persistence and recall will be the next meaningful differentiator, particularly for long-running agent tasks where models currently lose track of instructions.