Meta's Muse Wonder Debuts at $4/$20—Wonder Full Is API-Gated

Meta has released Muse Wonder and Muse Wonder Full, marking its first frontier-tier release after months of positioning Muse Spark as a lightweight, low-latency system. Trained at Meta’s gigawatt-scale Prometheus data center in central Ohio, the launch is the company’s first frontier release since Scale AI founder Alexandr Wang restructured Meta’s AI division in August of last year, as reported by TechCrunch. The operational split is immediate: standard Muse Wonder is available through the Meta Model API, while Muse Wonder Full is withheld from raw API access and restricted to Muse Code subscribers.

Architecture, Active Focus, and Test-Time Scaling

The new models move away from Muse Spark’s lightweight footprint toward larger, dense workloads built for sustained multi-step tasks. Meta describes two architectural behaviors behind this shift. Active Memory carries context across conversations and retrieves relevant details when earlier information becomes important, while Active Focus keeps a task moving through alternative approaches, blockers, and new information.

Meta evaluated the models on FocusBench and SWE-Marathon for long-running software engineering, and Harvey’s Legal Agent Benchmark for source-document legal work. In Meta’s own testing, as reported by TechCrunch, Muse Wonder and Muse Wonder Full achieve comparable benchmark scores to OpenAI’s GPT-6 Astra and Anthropic’s Claude Fable 5.1, and score higher on MemoryBench than Astra, Claude Fable 5.1, or Claude Opus. These are vendor-reported figures and have not been independently verified.

Meta also reports a test-time compute result: in a comparison of Muse Spark 1.3 and Muse Wonder Full, both models improve on unfamiliar tasks as inference-time compute increases. At a given in-distribution post-training benchmark score, however, Wonder Full shows an upward intercept shift—higher accuracy on un-targeted tasks. Meta attributes that generalization gap to larger, stronger base models, arguing that base parameter scale remains important for broad agentic capability.

Pricing and API Availability

Through the Meta Model API, Muse Wonder is priced at $4.00 per million input tokens and $20.00 per million output tokens, according to TechCrunch. AI Mastery’s pricing data places that at exactly four times the input price of Anthropic’s Claude 4.5 Haiku and more than triple the input price of Meta’s Muse Spark 1.1, while matching OpenAI’s GPT-5.6 Sol.

Muse Wonder rates come from Meta’s release data reported by TechCrunch; competitor pricing reflects AI Mastery’s token pricing dataset as of September 14, 2026.

AI Mastery analysis

The central tension is the bifurcated delivery mechanism. By locking Muse Wonder Full inside Muse Code, Meta is following the proprietary playbook of its closed-source peers: the most capable weights sit behind a vertical developer interface, while an intermediate model remains on the public API. As explored in our analysis of how token efficiency is repricing production AI, teams cannot treat model checkpoints as interchangeable commodities when their most capable variant cannot be integrated into custom orchestration pipelines. The practical choice is between the standard $4/$20 Wonder tier or Meta’s first-party IDE.

Active Memory and Active Focus are similarly worth scrutiny. Meta advertises persistence across days, weeks, and months, but production agents relying on vendor-managed active context often hit silent state degradation without deterministic memory boundaries. Until Meta documents the underlying primitives—test-time compute search, automated scratchpads, or managed retrieval—architects should treat multi-week continuity as an orchestration layer, not a raw model-weight guarantee.

The launch confirms Meta has shifted from a lightweight API-first model strategy toward monetized frontier infrastructure. With the full model gated behind Muse Code and the API tier positioned alongside GPT-5.6 Sol, Meta is betting that frontier agent capability justifies vertical integration and subscription lock-in.

Sources

• Primary: Muse Wonder: A Frontier Model For Personal & Professional AI Agents
• Reported by: TechCrunch AI: Meta releases its first frontier model, Muse Wonder
• Competitor prices: AI Mastery token calculator, figures as of 2026-09-14