Muse Spark 1.2

by Meta

A coding-focused update to Meta’s Muse Spark frontier model, pitched to challenge Claude Opus 4.8
and GPT-5.6 on coding benchmarks at low API cost. Powers Meta’s terminal coding agent, Muse Code.

Part of the Meta Muse family — the coding-tuned iteration of the original April 2026
Muse Spark launch.

Specs & positioning

  • Benchmarked by Meta and reviewers against GPT-5.6 (Data Max), Claude Opus 4.8 Max, and DeepSeek V4
    Pro on Artificial Analysis, DeepSWE, Terminal-Bench 2.1, and GDPval A v2
  • Cited as near-frontier on several benchmarks, at pricing well below Claude Opus 4.8 Max on
    OpenRouter
  • Powers Muse Code — a terminal-based coding agent (Mac/Linux) with persistent background agents,
    isolated sub-agents for large jobs, a replayable event log for crash-safe recovery, and commands like
    /pl (approval-gated planning), /grill (stress-test a plan), and /goal (long-horizon execution)

Timeline

DateEvent
2026-08-06Muse Code launches, powered by Muse Spark 1.2, pitched as a Claude Code competitor
2026-08-22Coverage frames Muse Spark 1.2 as directly challenging Opus 4.8 and GPT-5.6 on coding

Reception

Mixed. Coding/dashboard/landing-page demos (including inside OpenCode) are described as fast and
polished, with benchmark claims putting it close to or ahead of competing models in some evaluations.
Independent game-style/creative testing (RPG, flight sim, racing, combat demos) found it notably
weaker — buggy controls, poor camera behavior, weak visual polish — and rated it below Qwen 3.8, GPT-5.6
Sol, and Fable 5 on quality and presentation in that context. The gap suggests strong benchmark/coding
performance doesn’t transfer evenly to more open-ended generative tasks.