Meta ships Muse Code, and Opus 5 still beats it on paper

Meta Superintelligence Labs shipped Muse Code, a terminal coding agent running on Muse Spark 1.2 — its third model release in four months. Vendor-reported numbers: 59.3% on DeepSWE 1.1 and 82.9% on Terminal-Bench 2.1, with a 1M-token context window at $1.25/M input and $4.25/M output. macOS and Linux, one-line install, also on the Meta Model API and OpenRouter.

The framing matters more than the number. Meta's own comparison chart puts Claude Opus 5 at 86.7% on Terminal-Bench — ahead of Muse Spark 1.2, not behind it. Rivals in the deck were Grok 4.5, GPT-5.6 Terra, Gemini 3.6 Flash and Kimi K3, and The Register noted the scores cluster tightly, with Muse "close to the best, but never the very best." Beating Gemini Flash is a low bar to lead with.

And these are Meta's own harness results. Nobody — Muse Spark 1.2 or Opus 5 — has an independently verified Terminal-Bench 2.1 entry yet. When Muse Spark 1.1 was verified externally it came in at 76.2% against a published 80.0%: a 3.8-point haircut. Apply the same discount here and 82.9% looks a lot less like a headline.

The genuinely interesting bit isn't the benchmark at all — it's the runtime. Muse Code keeps persistent async background agents alive across a session and appends every model call, tool run and edit to a local event log, so work survives a crash. That's an engineering choice, not a leaderboard flex.

Why it matters: the coding-agent race is now won on runtime durability, not benchmark deltas — and vendor-reported scores keep shrinking under independent testing.