{
  "id": 186558,
  "title": "The 800 mistakes that could reshape Meta’s AI coding strategy",
  "url": "https://urgent.news/2026/08/05/the-800-mistakes-that-could-reshape-metas-ai-coding-strategy",
  "topic": "ai",
  "section": "AI",
  "published": "2026-08-05T18:06:05.000Z",
  "source": {
    "name": "The New Stack",
    "slug": "the-new-stack",
    "url": "https://thenewstack.io/meta-metacode-engineer-training/"
  },
  "original_language": "en",
  "account": "Meta is soliciting assistance from thousands of software engineers to help refine its internal AI coding tools by identifying and correcting mistakes. Last month, Meta's vice president of Applied AI Engineering, Maher Saba, instructed engineers to submit at least one \"diff\" per week using MetaCode, an internal AI coding agent. This process has already led to over 800 corrections and 7,000 weekly active users, with colored badges awarded to employees based on their contributions. MetaCode allows Meta to observe the entire process, from the AI's initial attempt to the engineer's correction and subsequent testing or review. As more submissions accumulate, Meta may identify recurring issues for future improvement. However, the specifics of how these corrections will influence post-training remain unclear. Meta has not disclosed whether every patch and code review will automatically become training data. Meanwhile, Meta introduced Muse Spark 1.1 on July 9, 2026, aiming for better coding capabilities than its predecessor. Despite being praised internally, Muse Spark 1.1 still lags behind the strongest models on public tests. The DeepSWE leaderboard, which evaluates coding agents on 113 software engineering tasks, shows Muse Spark 1.1 scoring 53% compared to GPT-5.5 at 67% and Claude Opus 4.8 at 59%. OpenAI and Anthropic have since surpassed these scores. It is crucial to note that improvements to Muse Spark 1.1 should not be directly compared to enhancements in MetaCode's performance unless Meta explicitly establishes this connection. Moreover, the rising cost of token usage as Meta scales its operations poses a significant challenge. With Muse Spark 1.1 costing $1.25 per million input tokens and $4.25 per million output tokens, the financial implications are substantial. The pricing becomes even more favorable when considering OpenAI's recent 80% reduction in GPT-5.6 Luna costs to $0.20 per million input tokens due to market pressures. This price war complicates Meta's position, as its model is currently cheaper than leading competitors, yet these rivals continue to decrease their prices rapidly. Other companies, like Alibaba, are also exploring ways to integrate AI-generated code with real-world engineering practices. Alibaba recently ran its Qwen model on a 16-day autonomous coding streak, publishing every commit to GitHub, demonstrating an alternative approach to using production coding as a training signal. As of now, Meta's primary value lies in the mistakes made by MetaCode during development, as these provide valuable insights for improvement. Meta continues to develop more coding and productivity tools, with MetaCode possibly becoming one of them, or remaining an internal tool.",
  "summary": "Meta is asking thousands of its software engineers to help train its internal AI coding tools by simply fixing code, The post The 800 mistakes that could reshape Meta’s AI coding strategy appeared first on The New Stack .",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 4,
    "also_reported_by": [
      {
        "outlet": "9to5Mac",
        "title": "Meta launches Muse Code AI coding agent for macOS and Linux",
        "url": "https://urgent.news/2026/08/05/meta-launches-muse-code-ai-coding-agent-for-macos-and-linux",
        "published": "2026-08-05T19:25:19.000Z"
      },
      {
        "outlet": "AppleInsider",
        "title": "Meta's Muse Code is yet another AI coding agent on macOS",
        "url": "https://urgent.news/2026/08/05/metas-muse-code-is-yet-another-ai-coding-agent-on-macos",
        "published": "2026-08-05T20:50:46.000Z"
      },
      {
        "outlet": "Tech in Asia",
        "title": "Meta launches Muse Code beta for AI coding",
        "url": "https://urgent.news/2026/08/06/meta-launches-muse-code-beta-for-ai-coding",
        "published": "2026-08-06T00:10:53.000Z"
      }
    ]
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}