Muse Wonder Full leads SWE-Marathon, but trails on terminal work
The Aggregate Digest — Tuesday, October 13, 2026
Muse Wonder Full leads SWE-Marathon, but trails on terminal work
Muse Wonder Full arrives with a small lead on SWE-Marathon’s software-engineering test. Its results on a separate command-line test are weaker, making the type of coding work an important part of the comparison.
Twelfth overall. Wonder Full enters at an Aggregate Rating of 1931, level with Claude Opus 5 and seven rating points behind Claude Sonnet 5.5. The smaller Muse Wonder ranks thirteenth on 1913.
The Aggregate, October 13, 2026. Selected neighbours; model ranks exclude human references and include private models.
A narrow engineering lead. On the primary SWE-Marathon v1.1 board, Wonder Full running in Muse Code at max reasoning passes 53.1% of tasks. Wonder in the same configuration scores 52.5%; Sonnet 5.5 at max scores 51.2%. These are results from the named agent configurations, so the comparison includes their tools and execution setup.
Primary SWE-Marathon v1.1 board, October 13, 2026. All rows use max reasoning; Wonder models run in Muse Code. Other source configurations remain distinct.
Terminal tasks tell a different story. Artificial Analysis’ Terminal-Bench 4.0 measures successful completion of command-line tasks. Wonder Full at max reasoning scores 57.1%, behind Sonnet 5.5’s max/default-fallback configuration on 63.6% and GPT-6 Astra at max on 59.1%. The smaller Wonder reaches 52.0% on this test.
Artificial Analysis, October 13, 2026. Exact reported reasoning and fallback settings retained; pass-at-one task success.
For repository work, that narrow engineering lead makes Wonder Full worth testing. Its terminal results give no reason to assume it will replace the stronger alternatives across coding tasks.
