Meta's Muse Wonder Crosses the Frontier Line — and Forces a Rewrite of the AI Capex Story

Muse Wonder Full lands at 56 on the Artificial Analysis Intelligence Index, two points behind Claude Opus 5.5 at 58, and ahead of GPT-6 Astra and Claude Fable 5.1 (both at 53). The smaller Muse Wonder scores 52, inside the same band as Fable 5.1 and Astra. The gap between Meta's best model and the best frontier model is now two points, not the10–15 that separated its previous flagships from the leaders.

Wonder Full is not uniformly best — it loses Terminal-Bench and HealthBench Pro to Opus 5.5 — but it sets the pace in three workloads that translate directly into enterprise pricing: legal review, software engineering, and long-running coding agents. On Harvey LAB-AA, Wonder Full scores 13.1 against Opus 5.5's 4.2, the largest single-benchmark gap any frontier model has posted against the leader on a high-value professional task.

Meta trained Wonder at the Prometheus data center in New Albany, Ohio — a 1-gigawatt cluster that Epoch AI puts at 536,000 H100-equivalents today, rising to 1.16 million H100-equivalents and 1,022 MW of IT power by Q3 2028 on a projected $38.7B total capital cost. Wonder runs on the same cluster Meta is using for everything else, which means the marginal cost of training it is mostly already in the capex line — a different cost structure from OpenAI or Anthropic, both of whom rent the bulk of their compute from Microsoft, Amazon, Google and CoreWeave.

Wonder is priced at $4 per million input tokens and $20 per million output tokens through the Meta Model API, with cached input at $0.20 per million tokens. That puts it roughly at parity with GPT-6 Sol ($2/$10) and Claude Opus 5.5, and well below GPT-6 Astra ($10/$50). Wonder Full, the larger model, is not available via API at all — only to Muse Code subscribers — which forces enterprise demand through Meta's own channel instead of letting third-party developers capture the gross margin.

The pattern is deliberate. By undercutting GPT-6 Astra by 60% on input and matching Opus 5.5 on price, Meta is positioning Wonder as the default budget tier of frontier models — a strategy that has historically worked for every computing layer below the leader, from cloud compute to EPYC server CPUs. The kicker is Wonder Full: by withholding it from the API, Meta keeps the highest-margin workload inside its own product stack rather than letting Azure or AWS resell it.

Q2 2026 was the quarter where Meta's free cash flow went negative on perception: revenue of $60.8B (+28% year-over-year) and net income of $15.85B were paired with $31.1B of capex and $784M of free cash flow, down roughly 91% year-over-year. The stock fell on the print because there was no model to point at. Today there is. The full-year 2026 capex guide, lifted to $130–145B in July from $115–135B in January and $125–145B in April, now has an output it produced: a frontier-class model trained entirely in-house on a cluster that was already in the capex base.

Anthropic told investors its annualized revenue run rate hit $65B at the end of July, up from $9B a year earlier, and is targeting a $2T Nasdaq listing in November 2026. That valuation rested on three claims: frontier-model parity with OpenAI, a runaway enterprise book, and a moat against open-weight competition. Muse Wonder takes a chip off each. On benchmarks, Meta is now inside two points of the leader — and ahead on legal and long-running coding, the workloads Claude is most associated with. On distribution, Meta is shipping the model inside an app that already tops app-store charts, alongside an installed base of three to four billion users across Facebook, Instagram, WhatsApp and Messenger that no pure-API startup can match.

Three upstream tiers just got their order book re-validated by a buyer with 3.7+ billion daily users. Nvidia shipped the Blackwell B300 silicon that runs Prometheus today; TSMC packages it through CoWoS (with Nvidia holding roughly 60% of 2026 CoWoS wafer starts and reportedly more than half of the 2026–2027 expansion); and SK Hynix and Micron supply the HBM4 stacks that feed the B300 — SK Hynix at an estimated 70% share of Nvidia's HBM4 allocation. Wonder does not change the demand picture, but it is the first hard confirmation that the 1-gigawatt Prometheus build is producing revenue-grade model output rather than waiting for one.