Meta claims Muse Wonder matches frontier AI, starts at $4/million tokens

By James Calloway

Meta will charge developers $4 per million input tokens for its new Muse Wonder model, the company confirmed this week as it rolled out two large language models (LLMs) it claims match the performance of leading frontier systems. Output tokens (the units of text an LLM generates) cost $20 per million.

The models — Muse Wonder and a higher-tier variant called Muse Wonder Full — add what Meta calls "active memory" and "active focus," features designed to keep an AI assistant on task across long sessions. Both are available now, Meta said.

What can the new models do?

In its announcement, Meta described Muse Wonder as "a major leap toward personal superintelligence" — language that echoes the company's stated long-term goal of building AI agents that match or exceed human performance on most cognitive work. The company framed the models as particularly capable on legal, financial, and health-related queries, and said they deliver "frontier coding performance," meaning code-generation quality on par with the strongest publicly available systems.

The two new capabilities do most of the storytelling. Active memory, Meta said, is "the ability to carry what matters to you across conversations and bring back the details that are relevant to the task at hand." Active focus, the company added, lets an agent "keep working toward your goal when a task gets hard." In practice, both target the same gap: most consumer chatbots (AI assistants that respond to typed prompts in a chat window) forget what users said ten messages ago, and most AI agents abandon multi-step work when an intermediate step fails.

How much will it cost?

The pricing splits cleanly between two audiences. Developers who plug into Meta's API (a programming interface that lets outside software send requests to a model and receive its outputs) pay $4 per million input tokens for Muse Wonder and $20 per million output tokens. Meta did not specify how those rates compare with the public prices of competing frontier models.

Muse Wonder Full, the higher-capacity variant, is not sold by the token. Meta bundles it into two subscription tiers of its coding product, Muse Code, at $50 per month for "power usage" and $200 per month for "maximum usage." The company did not disclose usage limits attached to either tier.

Meta said it will begin integrating the models into its personal agent, called Muse, starting with paid subscribers. The company did not say when the rollout will reach free users, or which countries will get access first.

How did Meta get here?

The launch caps a year in which Meta has poured resources into closing the gap it acknowledged falling behind. Chief Executive Mark Zuckerberg, in an interview with the Financial Times, attributed the progress to capital and science in roughly equal measure. "Getting to this point both required a large amount of compute and investment, and a lot of very rigorous scientific work," Zuckerberg said, crediting the company's superintelligence team.

Meta has said it expects to spend as much as $145 billion on AI this year. The figure dwarfs the company's reported AI investment in prior years and reflects the cost of training large models on thousands of specialized processors (GPUs, or graphics processing units, the chips that power most modern AI training runs). It also places Meta among the largest single corporate spenders on AI infrastructure, alongside Microsoft, Google, and Amazon.

The Muse-branded products themselves are new. Meta's prior flagship model line, the openly licensed Llama family, has not received a major release this cycle. That shift is itself a signal: Meta is now positioning its consumer and developer offerings under a different brand and a different sales pitch.

What is Meta not saying?

Meta's announcement does not include a single independent benchmark (a standardized test used to compare AI model performance) number. The company did not name a specific competitor whose product Muse Wonder matches, did not publish the model's parameter count (a rough measure of an LLM's size and capacity), and did not disclose the training data or training-compute total. The "frontier coding performance" claim, in particular, is unaccompanied by scores on any of the standard coding evaluations, such as SWE-bench (a benchmark that measures whether a model can resolve real GitHub issues) or HumanEval (a test of short, self-contained coding problems).

The omission is consistent with Meta's recent pattern. The company has historically published research papers for its Llama model family, but its newer Muse-branded products have shipped with marketing language rather than technical reports. That makes independent verification harder, and it gives prospective customers little to weigh beyond the company's own claims.

Takeaway: the strongest indicator of how good Muse Wonder actually is will be what paying developers and subscribers report over the next several weeks, not the launch announcement itself.

Why does this matter?

Two reasons. First, Meta's $145 billion AI budget is the clearest public signal yet of how much it now costs to compete at the top of the model market. If Muse Wonder is what Meta says it is, the company has bought its way back into a tier that OpenAI, Anthropic, and Google have dominated. If it is not, Meta has spent a fortune to reach parity with systems that are themselves about to be eclipsed by their makers' next releases.

Second, the "active memory" and "active focus" framing is not just a feature list. It is a bet that the next consumer AI fight will be over long-horizon tasks — the multi-step jobs that today's assistants still drop — rather than over raw chat quality. Meta is now selling that bet at $4 per million input tokens, and the price is the part of the announcement independent observers can actually verify.

Takeaway: Meta is asking customers and investors to take its word that Muse Wonder is a frontier model; the independent evidence will arrive in the benchmarks the company has not yet published.