Smithers Benchmarks

Frontier coding benchmarks run through a Claude-only delegation fleet.
delegation-chain (Opus plans/reviews, Opus->Opus, Opus->Sonnet implement)

Rows marked pending await the first full cloud pass. Public rows are each benchmark's own leaderboard. Small subsets are never comparable to full-benchmark numbers.

SWE-Bench Pro

ScaleAI/SWE-bench_Pro (test, 731)

731 multi-file tasks across 11 professional repos (Go/Python/JS/TS), hidden tests in canonical per-instance Docker.

ModelScaffoldPass@1 (%)nSubset / caveat
smithers (Opus + Sonnet) pending delegation-chain full-731 pendingClaude-only fleet run pending

RoadmapBench

UniPat-AI/RoadmapBench (115)

Long-horizon multi-target version upgrades (median ~3,700 LOC / ~51 files); the best single agent on the official board resolves under 40%.

ModelScaffoldResolved Rate (%)nSubset / caveat
public SOTA (best single agent) public official 39 115 full-115approximate published best
smithers (Opus + Sonnet) pending delegation-chain full-115 pendingClaude-only fleet run pending

Claw-Eval-Live

Claw-Eval-Live (105 tasks / 17 families)

Agentic enterprise workflows scored on observable execution evidence; no public model clears the 70% pass bar.

ModelScaffoldScore (pass bar 75)nSubset / caveat
Claude Opus 4.6 public official 83.6 105 full-105public leaderboard
GPT-5.4 public official 81.7 105 full-105public leaderboard
smithers (mixture) pending mixture-gateway pendingcannot run faithfully Claude-only (native OpenAI gather leg + neutral Gemini judge); asterisked variant only

SWE-EVO

SWE-EVO (48 release-sized tasks / 7 repos)

Carry a Python project across a whole release transition; frontier agents resolve only ~21-25% (full-48).

ModelScaffoldResolved Rate (%)nSubset / caveat
gpt-5.4 public official (OpenHands) 25 48 full-48Wilson 95% CI [14.9, 38.8]
gpt-5.2 public official (OpenHands) 18.75 48 full-48Wilson 95% CI [10.2, 31.9]
smithers (Opus + Sonnet) pending delegation-chain dvc subset pendingClaude-only fleet run pending; dvc subset is easier than full-48, not directly comparable