FULL STORY

Meta Launches Muse Code: Topping Benchmarks via Data-for-Pricing Strategy

Meta launched Muse Code and Muse Spark 1.2, employing a low-cost strategy that quickly topped multiple industry benchmarks for financial agents and cost-efficiency.

2026-08-06 ~ 2026-08-07 · 4 episodes · 65 posts

Episode 1 · Meta Launches Terminal Coding Agent Muse Code (2026-08-06, 44 posts)

Meta AI Research officially released Muse Code (beta), a terminal coding agent, along with its driving model Muse Spark 1.2. The tool is now available to users globally. Muse Code is designed to handle complex software engineering tasks within large codebases, marking a significant advancement for Meta in the coding agent space.

已确认

  • 产品与模型: Muse Code is currently in beta and is powered by Muse Spark 1.2, a brand-new model specifically optimized for coding capabilities.
  • 核心能力: The agent can handle end-to-end software engineering tasks across large codebases, including planning code changes, writing code, and verifying results.
  • 运行机制: Muse Code is equipped with multiple persistent asynchronous background agents, supporting the continuous processing of long-term, complex tasks in the background.

为什么重要

  • The release of Muse Code provides developers with a native terminal agent tool. Its asynchronous background and multi-file processing capabilities directly target complex engineering pain points in large, real-world codebases, promising to significantly boost the automation level of long-term software engineering projects.

24 more related posts →

Episode 2 · Meta Launches MuseSpark 1.2 and MuseCode, Disrupting Market with Data-for-Discount Pricing (2026-08-06, 15 posts)

Meta has unveiled MuseCode, its first terminal-based coding agent environment, alongside the accompanying large model MuseSpark 1.2, aiming to gain a foothold in the AI coding market. Its core strategy is "trading price for data"—using extremely low API pricing to acquire user data for model training, which has sparked industry-wide concerns over AI price wars and data privacy.

已确认

  • 产品发布：Meta introduced MuseCode (similar to Claude Code) and MuseSpark 1.2 (similar to Opus). MuseCode focuses on cross-file complex engineering tasks within large codebases, utilizing a main loop combined with resident asynchronous background agents, featuring precise replay and restart-safe capabilities.
  • 定价策略：If users allow their data to be used for training (the "contributor" option), API prices will drop significantly. The input price is as low as $0.10 per million Tokens, which is 12.5 times cheaper than the standard rate; cache and output prices are $0.002 and $0.20 per million Tokens, respectively.
  • 价格对比：@teortaxesTex and @歸藏的AI工具箱 pointed out that this cost is even lower than V4-Flash, and only about 70% of the price of competitors like DeepSeek.

为什么重要

  • 商业模式重塑：@ImaginaryDinner2710 and @teortaxesTex believe this indicates Meta is currently prioritizing the acquisition of high-quality coding data to improve its products over profit margins. This new model of trading ultra-low prices for data could disrupt the existing AI coding market landscape and trigger a new wave of price wars.

Episode 3 · Muse Spark 1.2 Tops Finance Agent Benchmark (2026-08-06, 4 posts)

Muse Spark 1.2 topped the Vals AI Finance Agent v2 benchmark by becoming the first model to exceed 60% accuracy, while maintaining extreme cost-efficiency at only $0.69 per test.

Episode 4 · Meta Muse Spark 1.2 Tops Cost-Efficiency Frontier, Costing 1/6 of Claude (2026-08-07, 2 posts)

According to Artificial Analysis, Meta's Muse Spark 1.2 tops the Pareto frontier of intelligence vs. cost, achieving a score close to Claude Opus at only one-sixth the cost, showcasing exceptional cost-efficiency.