Model results · Meta

Muse Wonder Full

Every published VulcanBench measurement of Muse Wonder Full (Meta), newest first: combined score, Code quality, runtime, tokens and an estimated cost at every effort level Muse Code offers in Contributor mode on Frontier v4, graded by deterministic hidden tests on real tasks. Measured through Muse Code 1.4.6 and 1.4.9 on a Muse Code subscription in Contributor mode, where Meta may train on submitted content, with Muse Code’s skill reminder off. Its Code quality is judged by Claude Opus 5.5 and GPT-6.1 Sol, a different panel from the other columns, because Muse Spark 1.3 is a Meta model. For the earlier Muse Spark 1.2, see its own page.

Muse Wonder Full across every effort level

In Muse Code across four effort levels: combined score 86.95 at Low, 88.27 at Medium, 88.19 at High and 87.59 at Extra-high (judged n=23, 23, 22, 22; 84.36 and 83.78 at High and Extra-high with one 3-hour timeout each counted as 0). Tasks passed 18, 21, 19 and 19 of 23; 27.3 to 36.2 minutes and an estimated $3.93 to $5.08 a task. Best level Medium, 23rd of 67 columns on the board.

October 17, 2026 · 23 tasks · 4 effort levels · 92 runs of Muse Wonder Full · VulcanBench Frontier v4 · Code quality protocol v3.26

Muse Wonder Full against Claude Opus 5.5, GPT-6.1 Sol and GPT-6 Astra: combined score, time and cost

Level with the frontier at Low and second at Medium, but no gain from more effort, while the slowest and most expensive of the four at about 27 to 36 minutes and an estimated $4 to $5 per task.

October 15, 2026, updated October 17 · VulcanBench Frontier v4 · Low to Extra-high

Analysis of traces comparing Muse Wonder Full and GPT-6.1 Sol at Low effort

27.2 minutes per task against 6.3, with about three times the model calls and ten times the output tokens, while each step took about the same time.

October 14, 2026 · VulcanBench Frontier v4 · Low effort traces

Redacted agent traces: Muse Wonder Full on Frontier v4

The trace repository for the sweep in the report above, published separately on GitHub, so you can read what the agent did on each task.

GitHub · vulcanbench-musewonder-full-traces

Next

How would Muse Wonder Full do on your stack?

Public results answer "which model is stronger on this suite." Whether Muse Wonder Full belongs in your routing table depends on your languages, your codebases, and your task mix: that gets measured, not guessed. Grading rules are in the methodology.

Optimize your model routing All reports