Benchmark leaderboards
Every public benchmark board ModelCap ingests, as its publisher reports it: human-preference Arena ratings, agentic and function-calling suites, coding pass rates and ARC-AGI reasoning. Each board has its own leaderboard listing every tracked model with a published result, the source's rank, and the model's ModelCap Index position, API price and context window.
Snapshot as of 17 August 2026
Coding
Arena's coding category scores the same blind pairwise votes restricted to programming prompts, so it reflects which models people preferred for code, not a fixed test set.
Claude Fable 5 leads 136 tracked models at 1554 · published 12 August 2026
Agents & tool use
LMArena's Agent board evaluates models running multi-step agentic sessions with tools; scores are the published win share with its official interval, per evaluated configuration.
Claude Opus 5 leads 37 tracked models at 12.2% · published 13 August 2026
Reasoning
ARC-AGI-2 measures fluid reasoning on novel visual grid puzzles that resist memorisation.
GPT-5.6 Sol leads 50 tracked models at 92.5% · published 13 August 2026
Reasoning
ARC-AGI-3 moves the ARC challenge into interactive game-like environments where the model must learn the rules while acting.
Claude Opus 5 leads 10 tracked models at 30.2% · published 13 August 2026
Coding
SWE-bench measures whether a model can resolve real GitHub issues by producing a patch that passes the repository's tests.
Gemini 3 Flash Preview leads 13 tracked models at 75.8% · published 19 February 2026
Agents & tool use
The Berkeley Function Calling Leaderboard scores how accurately a model calls tools across simple, parallel, multi-turn and agentic function-calling tasks; the overall score is out of 100.
GLM 4.6 leads 18 tracked models at 72.4 · published 12 April 2026
General capability
Artificial Analysis Intelligence Index is a public benchmark board published by Artificial Analysis.
Claude Opus 5 leads 155 tracked models at 63.1 · published 17 August 2026
7 public boards: Arena coding leaderboard; LMArena Agent leaderboard; ARC-AGI-2 leaderboard; ARC-AGI-3 leaderboard; SWE-bench leaderboard; BFCL V4 function-calling leaderboard; Artificial Analysis Intelligence Index leaderboard. Each has its own leaderboard page listing every tracked model with a published result.
Arena coding: Claude Fable 5 (1554). LMArena Agent: Claude Opus 5 (12.2%). ARC-AGI-2: GPT-5.6 Sol (92.5%). ARC-AGI-3: Claude Opus 5 (30.2%). SWE-bench: Gemini 3 Flash Preview (75.8%). BFCL V4: GLM 4.6 (72.4). Artificial Analysis Intelligence Index: Claude Opus 5 (63.1). Snapshot as of 17 August 2026.
Arena ratings come from blind pairwise human votes converted to a Bradley–Terry (Elo-style) rating, so they measure preference. SWE-bench, ARC-AGI and BFCL are fixed task suites with a pass rate or accuracy score, so they measure a specific capability. ModelCap shows each in its own units and never averages them into one number.
The ModelCap Index combines the public boards a model has an admitted result on, weighting by coverage and publishing an uncertainty interval; a model with a single board result carries a wider interval than one measured on several. The methodology page documents the sources and weights.
A model appears only when the source board publishes a result that matches a reviewed catalogue identity. Models the source has not evaluated, results that cannot be matched to an exact identity, and non-canonical variants are left off rather than guessed.
The data plane refreshes each source on its own cadence and seals the result into the dataset the site serves; every leaderboard page re-renders within a minute of a new dataset. Each row shows the publication date the source provides.