Meta released Muse Code, its new AI coding agent, in beta on August 5, 2026. The tool, powered by the Muse Spark 1.2 model, is designed to help developers write and debug code. But early benchmark results show it lags behind competitors from Anthropic and OpenAI.
Benchmark scores and a missing rival
On Terminal-Bench 2.1, a widely used coding benchmark, Muse Spark 1.2 scored 82.9%. That's below Anthropic's Claude Opus 5, which Meta reported at 86.7%. However, the independent Terminal-Bench leaderboard lists Opus 5 at 89.1% — a discrepancy Meta hasn't explained. OpenAI's Sol tops that same leaderboard at 89.5%, yet Meta omitted Sol from its comparison charts entirely.
On DeepSWE 1.1, another benchmark, Muse Spark 1.2 scored 59.3%, placing third behind Claude Opus 5 and others. Meta also published a custom benchmark based on 440 pull requests from its own repositories. There, Muse Spark 1.2 scored 70.6%, roughly 9 points behind Opus 5. Muse Spark 1.2 does not yet appear on the public Terminal-Bench leaderboard.
Pricing as a differentiator
Meta is positioning Muse Code as a cheaper alternative to rival coding agents. Its API pricing is $1.25 per million input tokens and $4.25 per million output tokens. A contributor tier costs more than 10 times less, though details on that tier are sparse. The company is betting that lower costs can win over developers who don't need top benchmark scores.
Financial context: revenue up, profit down, big spending
Meta's revenue rose 28% to $60.8 billion last quarter. But operating profit fell 8% to $18.8 billion, and the operating margin dropped from 43% to 31%. The company spent $31.08 billion on capital projects in the quarter and guided to up to $145 billion for the full year. Meanwhile, Meta is reportedly in talks to lease compute to Anthropic for up to $10 billion over two years — a deal that would put Meta in the unusual position of renting capacity to a direct AI competitor.
Meta's acquisition of Scale AI for $14.3 billion in June 2025 also looms large. Scale AI's founder, Alexandr Wang, now heads Meta Superintelligence Labs, the division behind Muse Code.
What Muse Code offers that rivals don't
Muse Code includes background agents and sub-agents — features that Claude Code lacks. These allow the tool to run multiple tasks in parallel or delegate subtasks to specialized agents. Whether that feature set can offset the performance gap remains an open question.
For now, developers can test Muse Code in beta. Meta hasn't announced a full release date or said when Muse Spark 1.2 will appear on the public Terminal-Bench leaderboard.
