Back
Meta launches Muse Code with $4.25-per-million output token pricing

Meta launches Muse Code with $4.25-per-million output token pricing

Meta's pricing push gained fresh attention after Muse Spark 1.2 reached the top five on the Vals Index, posting 71.88% accuracy at $0.69 per test and undercutting higher-cost rivals.

Fact Check
Both components of the claim are confirmed by primary/authoritative sources. CNBC's primary reporting confirms Meta launched Muse Code on Aug 5, 2026 with $4.25 per million output token pricing (and $1.25 input). The official Vals AI benchmark and model pages confirm Muse Spark 1.2 ranks 5th on the Vals Index at 71.88% accuracy with the lowest cost per test among the top five (~$0.69), undercutting higher-cost rivals such as Claude Fable 5 ($11.00). Benzinga independently corroborates these figures. The Vals AI page shows a rounded cost of $0.70/test, but the benchmark page and reporting cite $0.69, a trivial rounding difference.
Summary

Meta Platforms' aggressive AI pricing strategy drew renewed focus after Muse Spark 1.2 rose to fifth place on the Vals Index benchmark across finance and coding tasks. The model posted 71.88% accuracy with 630 seconds of latency at a cost of $0.69 per test, climbing four places from Muse Spark 1.1 and coming in well below competing systems including Moonshot's Kimi K3 at $2.34, Anthropic's Claude Fable 5 at $11.00, Claude Opus 5 at $8.54 and OpenAI's GPT-5.6 Sol at $7.46. Venture capitalist Chamath Palihapitiya called the approach "Tactical Game Theory: Meta Scorched Earth," arguing that lower-cost, high-accuracy models could become a stronger competitive moat as power and compute constraints tighten across the industry. He said Meta is now better positioned to pursue that strategy because hardware and energy bottlenecks are emerging. The benchmark adds market context to Meta's broader AI monetization push through Muse Code and Muse Spark 1.2. Meta's beta coding agent and model family already show a contributor tier that charges $0.10 per million input tokens, $0.002 for cache hits and $0.20 for output tokens when users let Meta use prompts and model outputs to train later models. Users who do not share data pay Meta's standard rates of $1.25 per million input tokens, $0.15 for cache hits and $4.25 for output tokens. Meta shares were down 10.80% year-to-date, up 1.01% over the last month and down 22.88% over the year; the stock closed 0.14% higher at $588.77 on Wednesday and rose 0.77% in premarket trading on Thursday.

Terms & Concepts
  • Vals Index: A benchmark that compares AI models across tasks such as finance and coding using measures including accuracy, latency and test cost.
  • output tokens: Units of text an AI model generates in response to a prompt.
  • cache hits: Requests served from previously processed input, which can be billed at a lower rate.