nharnessedLe runUnharnessed sur GitHubInstaller
Terminal-Bench 4.0

47 tâches sur 66. Tous les tours sont là.

Claude Opus 5.5 at max effort. Open a task to read what the model thought, the code it sent, what came back, and what the verifier said. Next to each task is what it took Claude Code with the same model.

71,2 %Score
5 993Tours
12,19 $Par tâche
nharnessedClaude CodeTâcheToursTempsCoûtToursTempsCoût
Moyenne par tâche911h1112,19 $981h2014,30 $
distributed-dedup632h2520,20 $1424h3042,76 $layout-config-recreation3011h2923,47 $3304h0344,99 $ctr-optimization1532h098,36 $2164h3924,92 $live-database-cutover1431h0818,61 $1782h0933,09 $biped-contact-dynamics555h379,93 $1272h3224,39 $vba-userform-port12353 min17,08 $1931h4130,46 $math-eval-grader10844 min11,24 $1311h2723,61 $retro-console-soc1051h2517,11 $1762h1428,92 $takens-embedding-lean8053 min13,15 $2341h0122,83 $rs-archive-clone19553 min15,84 $2081h5024,09 $fp8-rmsnorm-gemm8553 min11,71 $1351h2819,70 $payments-pipeline-fix1531h3121,68 $1842h4129,50 $photonic-waveguide-routing2141 min6,59 $311h5114,27 $cumulative-layout-shift1522h5324,45 $2053h0631,80 $interleaved-vigenere1031h128,72 $1141h4115,73 $ontology-kg-querying6737 min11,14 $9255 min17,48 $formal-crypto9339 min12,72 $1191h1218,24 $uefi-bootkit981h2510,28 $1121h0814,52 $sglang-qwen-burst2131h0923,31 $2211h3827,48 $cargo-flight-dispatch2319 min3,61 $3738 min7,63 $freight-dispatch-shift1131h0619,21 $1252h0022,84 $heat-pump-warranty7327 min7,53 $5446 min10,96 $react-lead-form6139 min9,04 $9251 min12,21 $music-harmony9944 min12,51 $891h0914,82 $wal-recovery-ordering4231 min6,13 $5346 min8,44 $legacy-utility-triage1421h209,31 $1791h0011,59 $embedding-drift-monitor6740 min6,57 $631h008,54 $fin-saccr-rwa3519 min4,15 $3235 min6,02 $risk-scorer-replay8335 min6,26 $8341 min7,64 $atrx-vep-crispr3814 min3,09 $3823 min4,44 $ks-solver-cpp7853 min10,50 $831h1511,63 $batched-eval-parity8535 min8,24 $7744 min9,05 $lake-temp-glm1703h2625,96 $1383h1726,58 $vllm-deepseek-streaming1231h0116,89 $1541h1717,48 $session-window-debug3338 min8,31 $3355 min8,74 $bun-sourcemap-leak3829 min6,24 $4934 min6,53 $production-planning *4129 min7,05 $3843 min7,31 $freecad-spring-clip3624 min4,98 $2526 min4,98 $shadow-relay265 min1,14 $178 min1,14 $nextjs-performance1201h0313,31 $1221h1613,16 $mp-checkpoint-consolidation4939 min3,56 $3624 min3,39 $cad-model5015 min2,82 $2918 min2,62 $hof-topology-interpenetration4017 min3,31 $2622 min2,96 $sound-change-cascade7347 min6,88 $621h026,51 $freecad-platform-drawing4715 min3,10 $4416 min2,68 $data-anonymization981h1115,10 $951h3814,66 $coq-block-bound1616 min2,93 $1717 min2,47 $medical-claims-processing9326 min9,78 $7036 min9,28 $intrastat-meldung9216 min5,67 $6718 min4,92 $layout-config-recreation28520 min3,05 $4721 min2,28 $roy-polymorph-cn3516 min2,65 $1215 min1,88 $foodstuff-beta-activity2837 min8,69 $2740 min7,68 $protein-autointerp-disulfide2829 min4,32 $728 min3,26 $mvcc-lsm-compaction2735 min4,51 $2431 min3,40 $freecad-impeller3543 min6,89 $3235 min5,64 $satb-audio-transcription8033 min4,91 $4725 min3,43 $pretrain-shard-corruption845h296,15 $5736 min4,45 $gsea-proteomics6634 min5,37 $4923 min3,29 $glycan-ms2-elucidation4330 min6,95 $2730 min4,59 $vpp-loss-divergence1371h0015,09 $11357 min12,21 $telecom-entity-resolution2291h0019,96 $1461h4016,78 $kv-live-surgery1141h1817,66 $7646 min10,90 $wdm-design1603h1227,28 $1573h0216,55 $html-js-filter1242h2733,84 $1311h1421,64 $vf2-speedup-networkx1623h2952,27 $1572h1232,90 $jax-speedrun-gpu *1616h4156,16 $1825h5432,83 $

* jax-speedrun-gpu : La première tentative est morte quand le sandbox a manqué de mémoire ; celle-ci a réussi.

* production-planning : Lancé après un correctif du system prompt : l’ancien demandait au modèle d’énoncer ses hypothèses, et il l’a pris au pied de la lettre, en supposant là où la tâche était explicite.