Meta Enters the AI Coding Wars With Muse Code

Meta has an AI coding agent now, and it arrived priced to start a fight. On 5 August the company's Superintelligence Labs released Muse Code, a terminal-based agent for writing and debugging software, together with Muse Spark 1.2, a coding-focused update to the model line. The pitch is aimed squarely at two products developers already use every day: Anthropic's Claude Code and OpenAI's Codex.

What makes Muse Code worth a second look is not the headline feature list but the architecture underneath it. The agent runs what Meta calls persistent background agents, which keep working through a session while the developer does something else, and it hands larger jobs to isolated sub-agents that run independently without stepping on each other. Every model call, tool use, approval and edit is written to a local, crash-safe event log. If the process dies partway through a long task, it can pick up exactly where it stopped rather than starting over.

A long-horizon demo, and a price to match

Meta backed the launch with a stress test rather than a slogan. In its own testing, the company said a single agent handled more than 1,000 tool calls across a 24-hour run while autonomously optimising GPU kernels on Nvidia Hopper hardware. That is the kind of unattended, long-running work most coding assistants still stumble over, and it is the specific claim rivals will want to reproduce.

The other weapon is cost. Muse Spark 1.2 runs at roughly $1.25 per million input tokens and $4.25 per million output tokens through Meta's API, below the going rate for comparable frontier models. Meta also offers a much cheaper contributor tier, dropping input pricing to as little as $0.10 to $0.30 per million tokens for developers willing to share data back. That is more than ten times cheaper than the standard rate, and it reads as a deliberate attempt to buy market share in a category OpenAI and Anthropic have owned.

On raw capability, Muse Spark 1.2 posted a score of 54 on the Artificial Analysis Intelligence Index at its highest reasoning setting, up from 51 for version 1.1 and 43 for the April release. The model is available through the Meta Model API and on OpenRouter.

Three big labs, one crowded terminal

The timing is not subtle. Coding is where large models have found their clearest commercial footing, and the tooling around it has become a proxy war between the major labs. OpenAI has been cutting prices across its GPT-5.6 family, and open-weight challengers such as DeepSeek's V4 Flash have been chasing the same agent benchmarks. Meta arriving with a genuinely different design and an aggressive price is less a surprise than a confirmation of where the money is going.

Whether developers switch is another matter. Claude Code and Codex have habits, integrations and trust built around them, and a cheaper agent with a clever event log does not automatically dislodge a workflow people already rely on. What Meta has bought itself is a seat at the table. The next few months of real-world use will decide if it keeps it.