VulcanBench VulcanBench

Reports

Long-form reports on how models work on VulcanBench tasks: what they read first, how they think, how they check their work and how that changes with effort.

Report01
How Claude Fable 5.2 thinks
It reads the reference binary before its first edit in all 115 runs, thinks in one long burst at Low, and spends its extra effort at Max on proving the rewrite. With GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5 for comparison.
2026-11-12 · Morgan Linton · VulcanBench Frontier v4, Low to Max effort · trace analysis
→
Read the benchmarks → Subscribe by RSS