Muse Code's 82.9% on Terminal-Bench 2.1: what a 4-point gap costs you in a real repo
Meta Superintelligence Labs shipped Muse Code, a terminal coding agent scoring 82.9% on Terminal-Bench 2.1 — here's what its 4-point gap to Claude Opus 5 actually costs.
Short version: Meta Superintelligence Labs' Muse Code is a credible terminal coding agent, and the reported 82.9% on Terminal-Bench 2.1 versus Claude Opus 5 at 86.7% is not the small gap it looks like. Flip the numbers — 17.1% failure versus 13.3% — and Muse Code fails roughly 29% more tasks than Opus 5 on the same suite. In an agent loop that chains ten tool calls before it reports back, that difference is where your afternoon goes.
The other thing worth staring at: Muse Spark 1.2 jumped +6.7 points on Terminal-Bench 2.1 and +6.3 on DeepSWE 1.1 over Spark 1.1. That is a large move for a point release, and large point-release moves on public agent benchmarks usually mean targeted post-training, not a broad capability jump.
Does the benchmark number transfer to your codebase?
Partially, and mostly for the shape of task the benchmark contains. Terminal-Bench 2.1 tasks are self-contained: a container, a goal, a verifiable end state. Your repo is not self-contained. It has a 400-line Makefile nobody understands, a test suite that needs a live Postgres, and three generations of internal conventions that exist only in the heads of people who left.
That is where a 4-point suite gap widens. Benchmark tasks have short chains. Real refactors have long chains, and per-step reliability compounds — a model that clears 95% of individual steps still only finishes a 15-step task about half the time. The agents that feel usable are the ones that notice they're off the rails and stop, not the ones that score two points higher on a leaderboard.
What actually breaks with terminal agents
The failure mode I keep hitting with every terminal agent, Muse Code's competitors included, is not bad code. It's bad state. The agent runs `git checkout -b`, edits six files, runs a test command that hangs on a missing env var, times out, then retries from a working tree it no longer models correctly. You end up reviewing a diff that is half fix and half archaeology.
So the questions I'd ask before wiring Muse Code into anything: does it operate on a scratch worktree or your live checkout, what's the token and wall-clock budget per task, and does it hard-stop on repeated test failures instead of grinding. Those three answers predict day-to-day usability better than either benchmark score.
Is the free contributor tier worth piping your codebase through?
The original launch report frames the contributor tier as the real strategic play here, and I think that read is right — for Meta. A free tier on a terminal agent is the cheapest possible way to buy agent trajectories: not just final diffs, but the full sequence of commands, failures, retries, and human corrections. That data is the scarce input for training the next Spark, and it is exactly what you cannot scrape off GitHub.
Which is a fair trade if you're on a side project. It is a different conversation if you're pointing it at a private monorepo under customer contracts. Read the data-retention terms as written, not as summarized in a launch post, and check whether opting out of training keeps the tier free. Meta acquihired Scale AI for a reported ~$14.3B largely for data operations; assume the data instinct carries over.
Should you use it?
Worth a real evaluation, not a switch. Give it three tasks you already know the answer to — one bug fix with a failing test, one dependency bump that breaks a build, one refactor across four files — and diff its output against whatever you use today. If Muse Code lands two of three, it's competitive. If it lands the first and thrashes on the other two, you found the 4 points.
I'd run it on open-source work and public repos immediately, because free is free and the trajectory trade is honest at that scale. I would not route a private production codebase through a brand-new tier from a lab whose parent company's core business is data, at least not until the retention language is boring.
FAQ
Is Muse Code better than Claude Opus 5 for coding?
On Terminal-Bench 2.1 as reported, no: Muse Code's 82.9% trails Claude Opus 5 at 86.7%. In failure-rate terms that's 17.1% versus 13.3%, so Opus 5 still finishes noticeably more tasks. Whether that matters depends on your task length — short, well-scoped jobs will feel nearly identical.
What is Muse Code built on?
Muse Code is a terminal-based coding agent from Meta Superintelligence Labs running on Muse Spark 1.2, which reportedly scores 82.9% on Terminal-Bench 2.1 and 59.3% on DeepSWE 1.1 — gains of 6.7 and 6.3 points over Spark 1.1.
What's the catch with the free contributor tier?
A free coding agent tier is typically paid for in data: command trajectories, failures, retries, and your corrections. Before pointing Muse Code's contributor tier at proprietary code, confirm the retention and training terms in writing and check whether opting out of training keeps the tier free.
Why is a 4-point benchmark gap a big deal?
Because agent tasks chain steps, and per-step reliability compounds. A 4-point gap on a single-task suite becomes a much larger completion gap on multi-step work — a model clearing 95% of individual steps finishes a 15-step task only about half the time.