Muse Code Review 2026

Muse Code is Meta's entry into the terminal coding agent race, announced by Mark Zuckerberg on August 5 and available immediately in public beta for macOS and Linux. It's powered by Muse Spark 1.2, a 1-million-token model co-trained with the agent itself, meaning the model was trained on the exact failure modes of agentic coding, not just code completion. The architecture is built around persistence: background agents that live for the whole session, and a local append-only event log that makes the runtime replay-exact, so a crashed session resumes precisely where it stopped.

The pricing is the real story. At $1.25/$4.25 per million tokens (standard) it already undercuts most rivals; the contributor tier at $0.10/$0.20 is roughly 12–21x cheaper. The catch (and Meta is explicit about it) is that contributor users consent to their prompts, completions, and repository context being used to train Meta's models, and the tier is only available in selected countries. For teams with proprietary code, the standard tier or enterprise zero-data-retention options are the realistic path.

On benchmarks, Muse Code scores 82.9% on Terminal-Bench 2.1, ahead of GPT-5.6 Terra but behind Claude Opus 5's 86.7%. Meta openly admits this gap and competes on price and architecture rather than raw capability. That's a healthy dynamic for the market: three real terminal agents now have verifiable numbers, and price competition is finally happening where it matters: per-token cost for long, credit-burning sessions. Pricing checked August 10, 2026.