magus v0.4.3 is out. See what's new
¶ View generated markdown
4 min read

Harness-effectiveness benchmark

Published 2026-09-10T15:28:56Z from results: 6 scored run(s), model claude-sonnet-5 (anthropic), magus v0.4.3-76-g63871b145 (63871b145) built 2026-09-09T22:55:47-04:00, host darwin/arm64/v8.

6 runs, 1 task(s), arms full and rampant, model(s) claude-sonnet-5, bootstrap seed 20260902.

Controls

Whether each task's check can tell a solution from its absence: the golden control applies the known solution and must pass, the null control touches nothing and must fail. A pass rate below is only worth reading where both hold.

task golden (pass/n) null (pass/n) checks discriminate
merge-config-falsy 1/1 0/1 yes

Headline: cost-of-pass

Expected dollars per correct solution (mean dollars / pass rate).

arm n passes pass rate Wilson 95% mean $ median $ cost-of-pass
full 3 3 100% [0.439, 1.000] $0.4441 $0.3996 $0.4441
rampant 3 3 100% [0.439, 1.000] $0.2196 $0.2120 $0.2196

Correctness against median tokens

The cost-accuracy frontier in text: pass rate beside the token spend it cost.

arm task n pass@1 median tokens median tokens (passes only) median $
full merge-config-falsy 3 100% 701714 701714 $0.3996
rampant merge-config-falsy 3 100% 386919 386919 $0.2120

Paired deltas, full minus rampant

Rep i of one arm is paired with rep i of the other. CI is a seeded 10000-sample bootstrap of the mean paired delta; p is Holm-adjusted across tasks. A verdict needs the CI to exclude zero AND the delta to reach 10% of the rampant median.

total billed tokens

task pairs median delta mean delta 95% CI relative p (Holm) verdict
merge-config-falsy 3 236268 355402.3 [200339.0, 629600.0] 92% 0.000 higher under full

dollars

task pairs median delta mean delta 95% CI relative p (Holm) verdict
merge-config-falsy 3 0.1385 0.2245 [0.1382, 0.3968] 106% 0.000 higher under full

wall clock (ms)

task pairs median delta mean delta 95% CI relative p (Holm) verdict
merge-config-falsy 3 35916 39241.7 [21857.0, 59952.0] 147% 0.000 higher under full

turns

task pairs median delta mean delta 95% CI relative p (Holm) verdict
merge-config-falsy 3 3 4.3 [1.0, 9.0] 39% 0.000 higher under full

tool calls

task pairs median delta mean delta 95% CI relative p (Holm) verdict
merge-config-falsy 3 3 4.3 [1.0, 9.0] 43% 0.000 higher under full

file reads

task pairs median delta mean delta 95% CI relative p (Holm) verdict
merge-config-falsy 3 0 -0.3 [-1.0, 0.0] -17% 0.585 inconclusive

re-read rate

task pairs median delta mean delta 95% CI relative p (Holm) verdict
merge-config-falsy 3 0.0000 0.0000 [0.0000, 0.0000] n/a 1.000 inconclusive

tool result bytes

task pairs median delta mean delta 95% CI relative p (Holm) verdict
merge-config-falsy 3 905 1364.3 [756.0, 2432.0] 35% 0.000 higher under full

pass@1 and pass^k

pass@1 is capability; pass^k (all k reps succeed) is reliability.

arm task k pass@1 Wilson 95% pass^k
full merge-config-falsy 3 100% [0.439, 1.000] 1
rampant merge-config-falsy 3 100% [0.439, 1.000] 1

Caveats

  • Reps per cell: full/merge-config-falsy n=3, rampant/merge-config-falsy n=3.
  • Under the 5-rep protocol: full/merge-config-falsy, rampant/merge-config-falsy. Treat those deltas as directional.
  • No activity trail for 3 run(s) (rampant-merge-config-falsy-r1-20260910T121032Z, rampant-merge-config-falsy-r2-20260910T121116Z, rampant-merge-config-falsy-r3-20260910T121152Z); their guard_events are null, not zero.
  • Dollars are the host's billed cost; the pricing table would have said 0.67x that, so the table is wrong for this model and only backs runs with no result record.
generatedbenchmarks/agent/reports/2026-09-10/anthropic/claude-sonnet-5.mdbenchmarksagentsharnessmeasurement
Last updated (0c2c811f)
Glossary

Op

A single tool-native command a target composes (long form: operation); the middle of the work hierarchy (Spell to Op to Target). See operations.

CI

An ordinary magusfile-defined target you compose yourself with magus\needs - magus does not hardcode its stages. Magus.RunCI treats it specially only in that it strips the rw charm, it is the anchor magus affected ci keys off, and a selected scope with no project declaring it is a load error rather than a silent no-op. See targets.

Run

One target executing under one magus invocation, such as magus run test web or magus affected ci. A run keeps its captured output behind an output reference. Every magus run is a run whether or not any job asked for it; see Job for how the two relate.

Conventions

This page uses none of the site's convention markers. The full set is on the conventions page.