Meta released Muse Code (beta), a terminal-based coding agent, on Wednesday. The launch puts Meta in direct competition with a fast-growing field of AI coding tools — including OpenAI's Codex, Google's Antigravity, Anthropic's Claude, and independent agents like Cursor and Devin — as developers increasingly delegate multi-file engineering work to autonomous models. Meta's own charts show the model outperforming OpenAI's Codex and Google's Antigravity on most coding tests, while trailing Anthropic's Claude Opus 5 on every benchmark Meta released.

Muse Code trails Opus 5 but adds a crash-safe log

Muse Code is powered by Muse Spark 1.2, an update to the coding model Meta opened to U.S. developers in July.

On the benchmarks Meta presented, Muse Spark 1.2 trails Opus 5 but beats Codex and Antigravity on most tests. Meta described the model as its “next step toward the frontier, with larger and much more capable models on the way.”

The company said Version 1.2 is better at code generation, debugging, and understanding large codebases after it increased the compute used for coding tasks during training.

In one case study, the model rewrote GPU kernels for NVIDIA Hopper chips through more than 1,000 tool calls over as long as 24 hours, working from a baseline it was instructed not to copy from existing libraries.

Muse Code keeps a local event log that records all model calls, tool runs, approvals, and edits. “This single source of truth makes the runtime replay-exact and restart-safe,” Meta wrote in its blog post about the log.

If a crash occurs, the agent can resume from where it stopped. The agent also keeps background subagents running for an entire session. That resilience matters for long-running tasks, where a single failure can otherwise cost hours of compute.

Mark Zuckerberg, Meta CEO, wrote in a post, “When a job is big enough, it fans out to separate sub-agents working in parallel in isolated worktrees.” He added, “Your working copy is never touched. In testing we had it build six features for a game simultaneously with no collisions.”

Releasing Muse Code in beta today. It’s a terminal coding agent that takes on complete software engineering tasks across large repos: planning changes, writing code, validating the results. Powered by Muse Spark 1.2, a coding-focused model update. pic.twitter.com/xqavk41w6v — Mark Zuckerberg (@finkd) August 5, 2026

/plan, /grill, and Meta's bet on cost

Muse Code includes several built-in commands. /plan turns a request into an approval-gated plan, /grill stress tests that plan until it holds up, and /goal drives toward completion of the stated objective.

Developers can install the agent on macOS or Linux with a single curl command. Muse Spark 1.2 is also available through the Meta Model API.

The model and the agent are designed to work together, with the Muse Code toolset keeping them compatible.

Some of the training data came from the older Muse Spark 1.1, which generated difficult coding problems and graded candidate answers. Meta said this loop helped the newer model follow instructions more closely.

Meta is telling investors it expects to spend between $125 billion and $145 billion this year on chips, data centers, and other infrastructure.

With Opus 5 out of reach on the benchmarks, the company is competing on price instead. That approach mirrors a broader pattern in the AI model market, where providers unable to claim the top benchmark spot are differentiating on cost, openness, or integration depth. A larger Meta AI model, codenamed Watermelon, is still in training.

If you're reading this, you're already ahead. Stay there with our newsletter.