Capability across Claude generations

ECCN commodity classification · Opus generation ladder · verified set (n=23) · METR-style trendline of accuracy vs. model release date

0.00.20.40.60.81.0Opus 4.1Aug 2025Opus 4.5Nov 2025Opus 4.6Feb 2026Opus 4.7Apr 2026Opus 4.8May 2026Model release date →Accuracy on verified set (n=23)SERIESgrade · no toolsexact · no tools
Tools condition pending. The agentic (CCL-navigation) run did not complete — the API account hit its credit balance partway through. Only the no-tools trendline is shown here. Re-run --agentic --run-id gen_agentic once credits are available, then rebuild with both --notools and --agentic.
GenerationReleasedexact · no toolsgrade · no tools exact · toolsgrade · toolserrors (tools/no)
Opus 4.12025-08-050.34780.4609———/0
Opus 4.52025-11-240.39130.5174———/0
Opus 4.62026-02-040.39130.5043———/0
Opus 4.72026-04-140.39130.487———/0
Opus 4.82026-05-280.21740.3957———/0
Capability snapshot, not an equalized comparison. Each generation runs in its strongest native reasoning config — Opus 4.1/4.5 use extended thinking (budget_tokens); 4.6/4.7/4.8 use adaptive thinking; 4.1 has no effort parameter. n=23 verified questions, so each item is ≈0.04 of the score — read the shape of the trend, not single-point gaps. "tools" = the model navigates the parsed CCL via the --agentic condition before answering.