Muse Wonder 1.0 test: GPT-6 Astra still leads on real bugs
On 105 real bugs that frontier models missed, GPT-6 Astra scored 45.0 to lead, while Muse Wonder 1.0 fell short of its marketing in hands-on testing.
So, I tested Muse Wonder 1.0 on real work. 2 repos. 105 bugs frontier models missed in Q1 2026. Find and fix what you can. The results: GPT-6 Astra (max): 45.0 Opus 5.5 (max): 41.7 Muse Spark 1.3 (max): 32.2 Muse Wonder 1.0 (max): 31.7 Grok 4.7 (xhigh): 29.0 I don't trust the maxed benchmarks. In my tests, it's similar to Spark 1.3, a bit slower, and 50% more expensive. Multiple runs. Next in this thread: Muse Spark Full 1.0. It's a separate model, available only via the subscription. ETA ~2.5 hours. Other effort levels for Muse Spark 1.0 can wait. P.S. And yes, Sonnet on max does crazy things making 7x turns than Astra.