Meta Enters the AI Coding War With Muse Code

Meta has made its most aggressive move yet into AI-assisted software development with the beta launch of Muse Code, a terminal-based coding agent, alongside Muse Spark 1.2, a coding-focused version of its proprietary frontier model.

Muse Code is designed to handle complete software-engineering tasks across large repositories, including planning changes, writing code, validating results, and running work through specialized AI agents. Meta is attempting to distinguish the system through persistent background agents, parallel development using isolated git worktrees, and a detailed local event log that allows interrupted jobs to resume without losing progress.

The underlying Muse Spark 1.2 model shows significant improvements over Muse Spark 1.1, although Meta’s own benchmark results still place Anthropic’s Opus 5 ahead across the three coding benchmarks highlighted in the article.

Perhaps the most consequential part of Meta’s strategy is pricing. Developers can use a dramatically cheaper Contributor tier, but in exchange they must allow Meta to use their prompts and completions for future model training. The standard tier costs substantially more but does not use customer prompts and completions for model training.

The launch also represents a major change from Meta’s earlier open-source AI positioning. Muse Code and Muse Spark 1.2 are proprietary, although Mark Zuckerberg suggested that Meta may have more to announce regarding open source in the future.

Key Points

• Meta is directly challenging the leaders in AI coding. Muse Code places Meta in competition with products such as Anthropic’s Claude Code and OpenAI’s Codex.
• Muse Code uses persistent background agents. Instead of repeatedly creating new helper agents for individual tasks, specialized agents can remain active throughout a session and retain knowledge about the repository.
• Large tasks can run in parallel. Muse Code can assign different pieces of work to sub-agents operating inside isolated git worktrees, preventing them from interfering with the developer’s working copy.
• Meta is emphasizing auditability and recovery. Model calls, tool operations, approvals, and edits are recorded in a local event log before execution. Meta says this makes the system capable of resuming long-running work after a crash.
• Muse Spark 1.2 was co-trained with Muse Code. Meta specifically optimized the model for the coding harness rather than treating the model and coding environment as completely separate products.
• The model has improved substantially. Muse Spark 1.2 gained 6.7 points over 1.1 on Terminal-Bench and 6.3 points on DeepSWE, although the article notes that some improvement may come from the new Muse Code harness.
• Anthropic still leads Meta’s published benchmark comparisons. Opus 5 finished first on Terminal-Bench 2.1, DeepSWE 1.1, and even Meta’s internal coding benchmark.
• Meta demonstrated long-running autonomous coding. In one case study, Muse Spark 1.2 performed more than 1,000 tool calls over as much as 24 hours while optimizing GPU kernels.
• Pricing is a major competitive weapon. The Contributor tier costs $0.10 per million input tokens and $0.20 per million output tokens, making it dramatically cheaper than Meta’s standard tier.
• The discount comes with a major condition: Contributor-tier users explicitly allow Meta to use their prompts and completions for training future models.
• The cheapest tier still requires billing details. VentureBeat’s testing found that developers need a payment method before Muse Code will perform work.
• Meta’s open-source direction is now uncertain. Unlike the Llama strategy that helped establish Meta as a major open-model player, Muse Code and Muse Spark are proprietary products.

Key Quotes

“It’s a terminal coding agent that takes on complete software engineering tasks across large repos.”

— Mark Zuckerberg

“It kept finding substantial improvements well beyond the initial exploration phase.”

— Mark Zuckerberg

“I’ll have more to share on that soon.”

— Mark Zuckerberg, responding to a question about open source

Implications

Meta is no longer standing on the sidelines of agentic coding. The article describes terminal-based coding agents as one of the fastest-growing areas of enterprise AI, previously dominated primarily by Anthropic and OpenAI. Muse Code gives Meta a serious product in that competition.

The battle may increasingly be about the model and the agent system together. Muse Spark 1.2 was trained specifically with Muse Code, reflecting the broader move toward optimizing models and the environments in which they operate as a single system.

Persistent agents could become an important differentiator. Meta’s architecture is designed to reduce repeated repository exploration and allow agents to continue working with accumulated context throughout a session.

Long-running autonomy is becoming a key competitive benchmark. Meta’s 24-hour GPU optimization experiment is intended to demonstrate that an AI coding agent can keep making useful progress rather than plateauing after its initial work.

Meta’s aggressive pricing introduces a significant data tradeoff. Individuals and experimental users can access the model extremely cheaply, but only by allowing their prompts and outputs to enter Meta’s training pipeline. Enterprises working with proprietary code may therefore prefer—or require—the more expensive standard tier.

Meta’s relationship with open source has fundamentally changed, at least for now. The company that once promoted downloadable Llama weights as the future of AI is entering the coding-agent market with a proprietary model and proprietary coding harness.

The central questions remain unresolved: whether Muse Spark 1.2 can match Claude- and GPT-class systems on real production repositories, whether developers will trust Meta with sensitive code, and whether the Contributor tier’s enormous discount will be attractive enough to overcome those concerns.

Source: https://venturebeat.com/orchestration/meta-enters-the-ai-coding-wars-with-muse-spark-1-2-and-muse-code-with-persistent-async-background-agents