SWE-bench Verified, live, four arms: 13% cheaper than no compaction, and the highest reward. See the numbers

Full results — headroom (hd-cache) (SWE-bench Verified, 50 tasks)

Live through the harness, claude-code agent on aws/claude-sonnet-5. Cache-aware billed input cost (fresh $2/M · cache-read $0.20/M · cache-write $2.50/M) + output $10/M, recomputed from each trial's token tiers. See REPRODUCE.md.

Totals

tasks scored solved rate total billed cost mean steps cache-hit agent wall (sum)
50 40 80% $30.30 35.1 98.0% 304 min

Per-task

task reward steps cache_read cache_write billed cost
astropy__astropy-12907 1 18 748,919 57,946 $0.337
astropy__astropy-14365 1 26 1,272,402 58,903 $0.476
astropy__astropy-8707 0 69 4,138,129 51,704 $1.107
django__django-11095 1 31 1,345,510 25,951 $0.407
django__django-11211 1 88 5,431,961 60,005 $1.474
django__django-11477 1 67 4,107,126 51,990 $1.150
django__django-11790 1 21 861,970 20,021 $0.266
django__django-12050 1 15 779,835 36,263 $0.268
django__django-12308 1 25 1,215,734 32,452 $0.366
django__django-12858 0 46 2,429,734 40,825 $0.827
django__django-13128 1 44 2,494,094 119,734 $1.042
django__django-13363 1 37 2,141,441 43,113 $0.637
django__django-13568 1 21 874,455 22,660 $0.270
django__django-13810 1 25 1,331,718 40,535 $0.456
django__django-14034 0 17 728,678 21,269 $0.276
django__django-14349 1 9 326,266 18,376 $0.128
django__django-14559 1 25 1,082,511 28,069 $0.331
django__django-14792 0 60 3,626,095 41,427 $1.067
django__django-15128 1 43 2,294,587 39,154 $0.657
django__django-15380 1 26 1,208,799 28,913 $0.369
django__django-15572 1 14 558,435 19,735 $0.189
django__django-15930 1 28 1,295,892 26,081 $0.399
django__django-16145 1 24 1,028,867 22,790 $0.312
django__django-16502 0 38 1,916,315 30,526 $0.604
django__django-16667 0 15 602,641 19,496 $0.191
django__django-17087 1 20 828,058 21,272 $0.253
matplotlib__matplotlib-22719 1 28 1,338,523 28,324 $0.407
matplotlib__matplotlib-24570 1 22 961,212 21,954 $0.364
matplotlib__matplotlib-25775 1 135 13,265,857 104,983 $3.485
psf__requests-1142 1 14 541,350 19,393 $0.190
pydata__xarray-3151 1 19 824,871 24,550 $0.270
pydata__xarray-4966 1 25 1,195,388 31,643 $0.380
pylint-dev__pylint-4551 1 45 2,801,827 52,055 $0.861
pytest-dev__pytest-10051 1 17 682,749 17,574 $0.213
pytest-dev__pytest-7205 1 48 2,502,813 39,914 $0.752
scikit-learn__scikit-learn-10844 1 10 363,542 15,746 $0.130
scikit-learn__scikit-learn-13328 1 21 866,537 20,972 $0.267
scikit-learn__scikit-learn-14894 1 20 828,462 20,742 $0.271
scikit-learn__scikit-learn-9288 1 30 1,476,615 31,687 $0.476
sphinx-doc__sphinx-7454 1 26 1,287,627 32,187 $0.406
sphinx-doc__sphinx-8120 1 74 4,375,724 52,043 $1.231
sphinx-doc__sphinx-8638 0 32 1,068,489 56,612 $0.543
sphinx-doc__sphinx-9602 0 56 3,376,186 45,137 $1.036
sympy__sympy-13031 1 42 1,924,725 30,000 $0.727
sympy__sympy-13877 1 16 649,900 49,201 $0.302
sympy__sympy-15599 1 45 2,261,362 36,297 $0.718
sympy__sympy-17318 1 75 4,015,695 38,471 $1.127
sympy__sympy-19495 1 45 2,311,024 33,835 $0.818
sympy__sympy-21379 0 26 1,164,192 23,205 $0.422
sympy__sympy-23413 0 32 1,645,502 33,325 $1.046