Ranks compare reported scores in the same published row; ties share a rank.
Terminal-Bench-Science 0.1. §8.5: Haiku 10 trials per task, task-clustered SE ±3.7 pts; 2.3% flagged trials fail without fallback. Claude Code --bare, max effort, no egress. Sol is quoted from OpenAI.
Section 8, pages 111–139 of the Haiku 5.5 system card, read 8 October 2026. Values come from exact prose, tables and labelled figures. Unless specified, Haiku uses adaptive thinking at max effort and five trials with evaluation-specific context up to 1M. Each row retains its subset, metric, effort and source notes; no score is assigned to an unspecified SWE-Bench Pro version. Terminal runs use safeguards without fallback and no internet egress. ProgramBench uses filtered tasks and hidden-test pass rate without a time limit. OSWorld partial and strict results use the official 82-task offline subset. CAD uses 0–1 voxel IoU. Healthcare raw and length-adjusted scores stay separate. Life-science results disable Haiku biology safeguards. Independently run figures quoted by Anthropic remain publisher reports here. Missing comparison values are not inferred from rounded charts.
www.anthropic.com/claude-haiku-5-5-system-card
Ranks compare reported scores in the same published row; ties share a rank.
Terminal-Bench Science 0.1. Argon uses a 6x verifier timeout; comparison figures are quoted from the author leaderboard.
Page 5 of Google DeepMind’s model evaluation, read 1 October 2026. Gemini runs use the API at the highest thinking setting, generally pass@1 without majority voting or parallel test-time compute; smaller benchmarks average multiple trials. Other models use their highest available published reasoning results. Google quotes Vals AI for knowledge-work and Vibe Code results, Zapier’s private-set AutomationBench leaderboard, Proximal for FrontierSWE v2, Surge AI for RiemannBench and no-tools Chartography, and the CWE-bench v1 leaderboard. Argon’s DeepSWE uses mini-swe; Terminal-Bench 4.0 is self-computed for Argon and quoted from the author board for other models. PostTrainBench v1.1 uses OpenCode with one H100 and a 10-hour budget for all models. Terminal-Bench Science uses a 6x verifier timeout for Argon. LABBench 2 uses terminal, bioinformatics tools, Python, R and internet access for all models. GraphWalks uses BFS F1 on two separate subsets: 650 items up to 128k tokens and 200 items from 256k to 1M. Agent’s Last Exam uses ALE-Claw, a 5-hour window and binary pass rate, with safety filters enabled for Argon. OSWorld 2.0 uses offline partial reward, the best of three single-attempt runs for Argon, safety filters, 1080p, 500 steps, Gemini CUA with parallel batch tools and compaction, screenshot-only observation and the 08.08 patch; Claude’s combined online/offline results are omitted. LVBench runs without tools, using 1 FPS for Gemini, 800 frames for Astra, 300 for Fable and 600 for Opus. Missing cells remain blank. These are the report’s published figures, preserved separately from this site’s independent source snapshots.
storage.googleapis.com/deepmind-media/gemini/gemini_4_argon_model_evaluation.pdf
Ranks compare reported scores in the same published row; ties share a rank.
OpenAI reports the maximum score at any effort level, in its own research environment or through its API. Claude and Gemini figures are taken by OpenAI from those vendors' published reports, so this table is a comparison between publication claims, not between runs.
openai.com/index/gpt-6-astra/
Ranks compare reported scores in the same published row; ties share a rank.
Anthropic runs its own harness and states that margin size is a weakening guide to real differences at this level. Two rows rely on other publishers: AutomationBench figures come from Zapier, and the GPT figures are as reported by OpenAI. The Claude Opus 5.5 row was run with production safeguards enabled, so some security, biology and frontier-development tasks were completed by a fallback model.
www.anthropic.com/news/claude-opus-5-5
Ranks compare reported scores in the same published row; ties share a rank.
Terminal-Bench Science 0.1. Maximum effort, where OpenAI reports the cost per task beside the score: $5.47 for GPT-6.1 Sol, $23.21 for Opus 5.5 with its fallbacks and $23.80 for Astra. Astra keeps the highest score of the models charted. Vals runs this benchmark independently on its own harness and scores the same models lower, so the two are not comparable.
OpenAI publishes this release as score-versus-cost charts at five reasoning efforts rather than as a table, and reads the top of each chart as the model's score, so every cell here is the highest figure OpenAI prints for that model and the effort it belongs to is in the row detail. OpenAI ran its own models in its research environment or through its API and quotes the Claude figures from those labs rather than re-running them, which makes the columns a comparison of publication claims rather than of runs. Three more charts on the same page are not carried here: GDP.pdf, the factuality error rate, and the four alignment charts, which are safety measurements rather than benchmark scores. Claude Opus 5.5 is charted with its production fallbacks and Claude Fable 5.1 with an Opus 5 fallback, as OpenAI labels them, and the column names are restated in the form this site uses. One number here repeats a figure the lab published earlier at a different value: OpenAI printed 72.6% for Astra on OSWorld 2.0 with its Astra table of 3 September, and the 29 September chart reads 73.5% at maximum effort, so both readings stand as published instead of one being reconciled into the other.
openai.com/index/introducing-gpt-6-1-sol/