The benchmark comparison deserves scrutiny. The product case rests on whether it holds up after hundreds of actions.
In June 2025, Meta put $14.3 billion into Scale AI and hired its founder, Alexandr Wang, to run its AI effort.
Last quarter, it spent more than $31 billion building AI infrastructure.
Today it released a command-line coding agent.
Muse Code is Meta’s first public attempt to turn that reset into something developers can actually judge.
It keeps agents alive on a repository, lets them build context over time, splits bigger jobs into separate worktrees, and records every command, edit and model call. Meta says one run made 1,000 tool calls over 24 hours while optimizing an NVIDIA Hopper kernel.
Its model, Muse Spark 1.2, scored 82.9% on Terminal-Bench 2.1. Opus 5 scored 86.7%.
That gap is close enough for the real question to change.
Can Meta’s agent keep making good decisions after the 200th tool call, when a repository has old assumptions, half-finished work and tests that fail for reasons nobody wrote down?
That is where coding agents stop being demos and start becoming software businesses.
