Meta is requiring thousands of its software engineers to submit at least one code correction each week through its internal AI coding assistant, MetaCode, according to a report from The New Stack. The initiative, outlined in an internal memo from Maher Saba, vice president of Meta's Applied AI Engineering organization, aims to improve the company's AI coding models by collecting real-world corrections from experienced developers. So far, 7,000 weekly active users have submitted more than 800 fixes, and Meta has added colored badges to employee profiles to incentivize participation.
The company is using these corrections to refine Muse Spark 1.1 and to post-train an upcoming model known internally as Watermelon, the memo reveals. MetaCode captures the full engineering workflow: the original task, MetaCode's initial response, the engineer's correction, and any tests or reviews required for approval. Muse Spark 1.1, introduced on July 9, 2026, scored 53% on the DeepSWE 1.1 leaderboard, which evaluates coding agents on 113 extended software engineering tasks. At launch, Meta compared this with GPT-5.5 at 67% and Claude Opus 4.8 at 59%, though both OpenAI and Anthropic have since advanced: GPT-5.6 Sol reached 73% in mid-July, while Claude Opus 5 hit 74% after its July 24 release. Muse Spark 1.1 costs $1.25 per million input tokens and $4.25 per million output tokens through Meta's API, substantially less than what Anthropic and OpenAI charge for their leading models.
The report notes that public code repositories contain vast amounts of functional software but rarely preserve the model's first attempt, where it failed, and what an engineer modified before merging. MetaCode gives Meta visibility into that complete process, enabling the company to identify patterns in recurring mistakes during normal development work rather than through synthetic coding exercises. The memo describes a structured system in which engineers submit corrections after MetaCode produces flawed output, though Meta hasn't disclosed how it prepares or weighs these corrections during post-training, nor confirmed that every patch automatically becomes training data.
Cost considerations make this strategy particularly important for Meta. Coding agents consume tokens as they search repositories, test solutions, and revise code, and when multiplied across thousands of engineers running agents on lengthy tasks, expenses become substantial. If Meta can close the performance gap with competitors, it could operate coding agents company-wide without paying rivals each time an engineer assigns a task. However, the pricing landscape is shifting rapidly: OpenAI cut GPT-5.6 Luna costs by 80% in late July, dropping input tokens to $0.20 per million, driven partly by competition from efficient Chinese open-weight models. During Meta's second-quarter earnings call, CEO Mark Zuckerberg said the company has additional coding and productivity tools planned, though whether MetaCode will become a public product or remain internal is unclear. Meta isn't alone in this approach—Alibaba recently ran its Qwen model on a 16-day autonomous coding project with every commit published to GitHub, representing a different path toward the same goal of using production coding as a training signal rather than synthetic benchmarks. The competitive pressure is intensifying, with frontier models not only outperforming Muse Spark but also dropping prices at a pace that narrows Meta's cost advantage even as the performance gap persists. If training on internal mistakes proves more effective than relying solely on public code, Meta may gain leverage that extends beyond any single leaderboard score.