[1]
Arena Text Leaderboard
Measures: Human preference in anonymous model-vs-model battles across open-ended text tasks.
Use it for: General assistants and model-family context
Do not miss: A product such as ChatGPT or Claude can change its default model; Arena ranks specific model versions, not the whole product. Checked 2026-09-14 and re-checked 2026-09-21, 2026-09-28, 2026-10-05, 2026-10-12, 2026-10-19, 2026-10-26 and 2026-11-09: /leaderboard/text 308-redirects to /leaderboard/chat/text, which answers a Cloudflare managed challenge to plain HTTP clients and to a bare headless-Chrome DOM dump, so this check read the board through a persistent Chrome profile (Playwright driving the installed Chrome channel) and cross-read the rendered rows against the page's own embedded per-model JSON (rank, rating, vote count and modelKey for all 420 listings). The board served a new snapshot dated 2026-11-08 (9,091,367 votes, 420 models), replacing the 2026-10-22 snapshot (8,946,356 votes, 418 models). Every tracked row was re-read from it and all six keep the same exact model key as before; no rating moved at integer rounding, and vote counts fell between 141 and 684 on a board whose total rose by 145,011. Newer same-board listings were again seen and deliberately NOT added as rows: grok-4.5-text (#53, 1466 +/-4, 39,169 votes), grok-4.1-thinking (#54, 1465 +/-3, 66,085 votes) and deepseek-v4-pro-high-20260813-text (#58, 1464 +/-7, 9,593 votes). This registry carries one representative exact listing per product family, and the site's generators key rows by tool_slug, so a second row for the same family would silently overwrite the first in every generated comparison rather than adding evidence.
Open source ↗
[2]
Arena-Rank methodology
Measures: Bradley-Terry ratings with reweighting and confidence intervals from pairwise preference data.
Use it for: Explaining how Arena scores are produced
Do not miss: Human preference can reward style and verbosity; it is not an objective factuality test.
Open source ↗
[3]
SWE-bench Verified
Measures: Percentage of 500 human-validated real GitHub issues resolved by a coding model or agent.
Use it for: Coding models and coding-agent context
Do not miss: Scores depend heavily on the agent harness. mini-SWE-agent v1.x and v2.x results are not directly comparable. Checked 2026-09-21 and re-checked 2026-09-28, 2026-10-05, 2026-10-12, 2026-10-19, 2026-10-26 and 2026-11-09: the site still publishes a machine-readable results payload on the home page (an embedded leaderboard-data JSON block covering Multilingual, Test, Verified, Lite and Multimodal). The Verified group is unchanged at 180 rows dated 2023-10-10 to 2026-02-26; the newest is still a mini-SWE-agent 2.0.0 run on Gemini 3 Pro at 69.6% resolved, and the highest score in the group is Sonar Foundation Agent + Claude 4.5 Opus at 79.2% (2025-12-05), tied by live-SWE-agent + Claude 4.5 Opus medium at 79.2% (2025-12-15). No Cursor, GitHub Copilot, Replit, Windsurf or Devin product appears in that Verified set - the only Codex occurrences are model names inside other agents' rows - so no product-level coding row is added from it. The page text still confirms the 500-instance human-validated subset.
Open source ↗
[4]
LiveBench
Measures: Objective, frequently refreshed tasks across reasoning, math, coding, language, instruction following and data analysis.
Use it for: Contamination-limited general-model context
Do not miss: Always state the release; questions and model configurations change between snapshots. Checked 2026-09-21 and re-checked 2026-09-28, 2026-10-05, 2026-10-12, 2026-10-19, 2026-10-26 and 2026-11-09: the repository is live and active (latest commit still 2026-10-18, adding a deepseek-v4.1-flash-1017-max configuration), but its own changelog.md still ends at 2026-01-08 and the README still shows "Top models as of 30th September 2024". The material content sits on the website, which is branded "LiveBench 2.0 - a challenging, contamination-free long-horizon AI benchmark" and whose bundled release list still ends at two 2.0 releases: 2026-10-14 (/data/20261014/table.csv, 55 model rows on this check, up from 53 on 2026-10-26) and 2026-10-25 (/data/20261025/table.csv, 32 model rows, up from 30), both across the same seven categories (long_agentic, reasoning, coding, browser, data_analysis, research, design); no release dated after 2026-10-25 exists (probed 2026-11-01, 2026-11-07, 2026-11-08 and 2026-11-09 table paths, all 404). Generation v1 releases run 2024-06-24 through 2026-06-25; the v1 June table at /v1/table_2026_06_25.csv still holds 71 model rows across 23 task columns. The two generations use different task sets and must never be pooled or compared, and neither is a product-level ranking: treat LiveBench as release-pinned model evidence only.
Open source ↗
[5]
Artificial Analysis methodology
Measures: Independent intelligence, latency, speed, price, speech, image and video model evaluations.
Use it for: Cost/speed context and non-text modality methodology
Do not miss: API model measurements do not equal the complete end-user product experience. Re-checked 2026-11-09: the methodology page is live (HTTP 200, 373,607 bytes) and still serves no dated per-model score table, so no Artificial Analysis row exists in this registry.
Open source ↗
[6]
Terminal-Bench 2.1 leaderboard (frozen; subdomain now serves 4.0)
Measures: Accuracy of an agent + model + reasoning-effort configuration on 89 complex containerized terminal tasks.
Use it for: Coding agents and terminal-capable development tools
Do not miss: Checked 2026-09-14, re-checked 2026-09-21, 2026-09-28, 2026-10-05, 2026-10-12, 2026-10-19, 2026-10-26 and 2026-11-09: https://www.tbench.ai/leaderboard/terminal-bench/2.1 still returns a 308 redirect to https://www.tbench.ai/?version=2.1, and every route on that host still serves the Terminal-Bench 4.0 board (leaderboard slug "4-0-0"). The 4.0 board object still carries updated_at 2026-10-23T00:29:02Z, unchanged from the previous check, and still parses to 35 rows at 330 trials each with zero Cursor, Copilot, Windsurf or Replit entries - the rows cover only Claude Code, Codex, Grok Build, Muse Code and mini-SWE-agent configurations, with Claude Code + Opus 5.5 at max effort on top at 64.85% +/-3.11. No 2.1 rows can be read from this URL on any date, and no archived 2.1 board exists on the site, so nothing here can move the registry's coding row. Never present this as a model-only score; keep agent, exact model, effort, trial count and integrity metadata visible.
Open source ↗
[7]
Artificial Analysis Image methodology
Measures: Blind pairwise image preference with exact model/version and matched prompt conditions.
Use it for: Text-to-image and image-editing model context
Do not miss: Keep text-to-image and editing pools separate; old model versions are historical evidence only. Re-checked 2026-11-09: the page now resolves at https://artificialanalysis.ai/methodology/image (438,506 bytes), keeps separate text-to-image and editing sections, and serves no dated score table.
Open source ↗
[8]
Artificial Analysis Video methodology
Measures: Modality-separated video quality, latency and price comparisons.
Use it for: Text-to-video and image-to-video model context
Do not miss: Do not compare Elo across different video modalities or output settings. Re-checked 2026-11-09: the page now resolves at https://artificialanalysis.ai/methodology/video (488,436 bytes) and still states that each modality has its own Elo pool in the Video Arena, so silent and audio-enabled outputs are not compared directly and editing is not compared with generation. Its only table is a settings/parameters table; no dated score table is served.
Open source ↗
[9]
Artificial Analysis TTS methodology
Measures: Provider-voice and controlled-voice human-preference comparisons for text-to-speech.
Use it for: Voice-generation model context
Do not miss: Controlled-voice and provider-voice arenas answer different questions and must remain separate. Re-checked 2026-11-09: the page now resolves at https://artificialanalysis.ai/methodology/text-to-speech (512,850 bytes) and separates the provider-voice and controlled-voice sections. No dated score table is served.
Open source ↗
[10]
Terminal-Bench: Cursor CLI + Grok 4.5 submission
Measures: Maintainer-verified Terminal-Bench 2.1 submission record for one exact Cursor CLI configuration.
Use it for: A configuration-specific Cursor benchmark snapshot
Do not miss: Checked 2026-09-14, re-checked 2026-09-21, re-hashed 2026-09-28, 2026-10-05, 2026-10-12, 2026-10-19, 2026-10-26 and 2026-11-09 against the raw submission file (sha256 a73454861c91299b501637e6df6082f2f16cf9c9f6c98f07bce315189f82eb23, 29,033 bytes, byte-identical on every check): Cursor CLI 2026.07.08-0c04a8a + cursor/grok-4.5 at high effort, 79.33% +/- 1.46, 445 trials, 8.99% reward-hack disqualifications, dated 2026-07-09. Two limits sit on it: the Terminal-Bench 2.1 README states "Community submissions are currently closed for Terminal-Bench 2.1", so the 2.1 board is frozen, and the live 4.0 board carries no Cursor entry, so there is no current Cursor reading to compare it to. Scores from 2.1 and 4.0 are different harness versions and must never be compared with each other.
Open source ↗
[11]
Harbor Hub — Terminal-Bench 2.1 dataset
Measures: The live dataset record for Terminal-Bench 2.1: which tasks the frozen board ran and which version is published.
Use it for: Confirming the task set behind a Terminal-Bench 2.1 configuration score
Do not miss: Checked 2026-09-14, re-checked 2026-09-21, 2026-09-28, 2026-10-05, 2026-10-12, 2026-10-19, 2026-10-26 and 2026-11-09: the dataset page describes "Version 2.1 of Terminal-Bench" and still serves revision 6 as its latest published revision. The Terminal-Bench 2.1 repository independently holds 89 task directories (91 entries minus README.md and dataset.toml), matching the 89 tasks this page reports. Terminal-Bench 4.0 is not a "larger" task set: the Terminal-Bench repository holds 66 task directories at its v4.0.0 release tag (published 2026-08-26). 2.1 and 4.0 are different task sets, not a small/large pair. This page publishes the task list, not leaderboard scores; it cannot upgrade a 2.1 score into a 4.0 score.
Open source ↗