Terminal-Bench 4.0

47 of 66 tasks. Every turn is here.

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, on average over its five official runs.

The list runs from the task where the harness saved the most to the ones where it cost more. A time is how long the run took, without the waits on our subscription’s quota.

71.2%Score

5,993Turns

$12.19Per task

nharnessedClaude CodeTaskTurnsTimeCostTurnsTimeCost

Average per task911h11$12.19981h20$14.30

distributed-dedup632h25$20.201424h30$42.76 layout-config-recreation3011h29$23.473304h03$44.99 ctr-optimization1532h09$8.362164h39$24.92 live-database-cutover1431h08$18.611782h09$33.09 biped-contact-dynamics555h37$9.931272h32$24.39 vba-userform-port12353m$17.081931h41$30.46 math-eval-grader10844m$11.241311h27$23.61 retro-console-soc1051h25$17.111762h14$28.92 takens-embedding-lean8053m$13.152341h01$22.83 rs-archive-clone19553m$15.842081h50$24.09 fp8-rmsnorm-gemm8553m$11.711351h28$19.70 payments-pipeline-fix1531h31$21.681842h41$29.50 photonic-waveguide-routing2141m$6.59311h51$14.27 cumulative-layout-shift1522h53$24.452053h06$31.80 interleaved-vigenere1031h12$8.721141h41$15.73 ontology-kg-querying6737m$11.149255m$17.48 formal-crypto9339m$12.721191h12$18.24 uefi-bootkit981h25$10.281121h08$14.52 sglang-qwen-burst2131h09$23.312211h38$27.48 cargo-flight-dispatch2319m$3.613738m$7.63 freight-dispatch-shift1131h06$19.211252h00$22.84 heat-pump-warranty7327m$7.535446m$10.96 react-lead-form6139m$9.049251m$12.21 music-harmony9944m$12.51891h09$14.82 wal-recovery-ordering4231m$6.135346m$8.44 legacy-utility-triage1421h20$9.311791h00$11.59 embedding-drift-monitor6740m$6.57631h00$8.54 fin-saccr-rwa3519m$4.153235m$6.02 risk-scorer-replay8335m$6.268341m$7.64 atrx-vep-crispr3814m$3.093823m$4.44 ks-solver-cpp7853m$10.50831h15$11.63 batched-eval-parity8535m$8.247744m$9.05 lake-temp-glm1703h26$25.961383h17$26.58 vllm-deepseek-streaming1231h01$16.891541h17$17.48 session-window-debug3338m$8.313355m$8.74 bun-sourcemap-leak3829m$6.244934m$6.53 production-planning *4129m$7.053843m$7.31 freecad-spring-clip3624m$4.982526m$4.98 shadow-relay265m$1.14178m$1.14 nextjs-performance1201h03$13.311221h16$13.16 mp-checkpoint-consolidation4939m$3.563624m$3.39 cad-model5015m$2.822918m$2.62 hof-topology-interpenetration4017m$3.312622m$2.96 sound-change-cascade7347m$6.88621h02$6.51 freecad-platform-drawing4715m$3.104416m$2.68 data-anonymization981h11$15.10951h38$14.66 coq-block-bound1616m$2.931717m$2.47 medical-claims-processing9326m$9.787036m$9.28 intrastat-meldung9216m$5.676718m$4.92 layout-config-recreation28520m$3.054721m$2.28 roy-polymorph-cn3516m$2.651215m$1.88 foodstuff-beta-activity2837m$8.692740m$7.68 protein-autointerp-disulfide2829m$4.32728m$3.26 mvcc-lsm-compaction2735m$4.512431m$3.40 freecad-impeller3543m$6.893235m$5.64 satb-audio-transcription8033m$4.914725m$3.43 pretrain-shard-corruption845h29$6.155736m$4.45 gsea-proteomics6634m$5.374923m$3.29 glycan-ms2-elucidation4330m$6.952730m$4.59 vpp-loss-divergence1371h00$15.0911357m$12.21 telecom-entity-resolution2291h00$19.961461h40$16.78 kv-live-surgery1141h18$17.667646m$10.90 wdm-design1603h12$27.281573h02$16.55 html-js-filter1242h27$33.841311h14$21.64 vf2-speedup-networkx1623h29$52.271572h12$32.90 jax-speedrun-gpu *1616h41$56.161825h54$32.83

* jax-speedrun-gpu: The first attempt died when the sandbox ran out of memory; this run passed.

* production-planning: Ran after a fix to the system prompt: the old one asked the model to state its assumptions, and it took that literally, assuming where the task was explicit.