Independent analysis of AI models
Rankings from 297 public and community benchmarks. Every number links to where it was published, and the index is a JSON file anyone can change by pull request.
Intelligence
Open Index
Equal-weight average of 10 independent benchmarks: Terminal-Bench 4.0, ProgramBench v1 Fully Resolved, DeepSWE v1.1, ARC-AGI-2, SimpleBench, BullshitBench v2, FrontierMath Tiers 1–3 (v2), Finance Agent v2, AutomationBench (Zapier), GPQA Diamond (Vals) · Missing results may use peer estimates · Higher is better · Methodology
Capability Indexes
How models perform in specific domains. Each index averages the best-covered benchmarks in its category.
Open Coding Index
Incorporates 6 evaluations: Terminal-Bench Hard-AA (17%), Terminal-Bench v2.1-AA (17%), Arena WebDev (17%), Vibe Code Bench v1.1 (17%), ALE-Bench (17%), LiveCodeBench (Vals) (17%) · Higher is better
Benchmarks
Every benchmark with results for at least eight models released in the last year, including 1 from the community. Follows the filters above. Add yours.
CritPt
Artificial Analysis · Reasoning · 70 recent models
Humanity's Last Exam
Artificial Analysis · Reasoning · 70 recent models
AA-Omniscience
Artificial Analysis · Knowledge · 70 recent models
AA-Omniscience Accuracy
Artificial Analysis · Knowledge · 70 recent models
AA-Omniscience Non-hallucination
Artificial Analysis · Knowledge · 70 recent models
AA-LCR v1.1
Artificial Analysis · Long Context · 70 recent models
AA Intelligence Index v4.3.2
Artificial Analysis · Composite · 70 recent models
GPQA Diamond
Artificial Analysis · Knowledge · 48 recent models
GDPval-AA v2.1
Artificial Analysis · Professional · 68 recent models
Terminal-Bench v2.1-AA
Artificial Analysis · Coding · 49 recent models
AA-Briefcase v1.1
Artificial Analysis · Agentic · 64 recent models
SciCode
Artificial Analysis · Coding · 69 recent models
Open Weights
5 of the 29 ranked models with known licensing publish their weights. The best open model scores 62 (GLM 5.3) against 69 (Opus 5.5) for the best proprietary one.
Open Index: open-weights models
Open weights per Artificial Analysis model metadata · best effort per model
Output Tokens
Tokens generated per Intelligence Index task, including reasoning. More tokens means slower and costlier answers.
Cost
Measured cost to complete a task, not token price. Reasoning-heavy models can cost more per task even at lower list prices.
Speed & Latency
How long a task takes is on the leaderboard above (Time per Task). Raw token rates and first-token latency are shown here for reference; a fast token rate does not mean a fast answer once reasoning is counted.
Providers
How many API providers serve each model on OpenRouter, and how reliably.
Trends
Open Index score by release date, every reasoning effort. Turn on the frontier in the chart settings to trace the best model at each point.
All Models
Best reasoning effort per model, for the models and window picked above. Measured is the share of index weight backed by real results; the rest is estimated from close peers.
| # | Model | Open Index | Cost / task | Time / task |
|---|---|---|---|---|
| 1 | Opus 5.5max | 69.0 | $0.091 | 13 min |
| 2 | GPT-6 Astraxhigh | 66.8 | $0.198 | 5.3 min |
| 3 | Sonnet 5.5max | 64.9 | $0.083 | 15 min |
| 4 | Fable 5.1high | 64.5 | $0.610 | 6.5 min |
| 5 | GPT-6.1 Solmax | 64.4 | $0.085 | 10 min |
| 6 | Opus 5max | 63.5 | $0.102 | 13 min |
| 7 | GLM 5.3minimal | 62.3 | – | – |
| 8 | GPT-6 Solmax | 60.1 | $0.065 | 4.7 min |
| 9 | Gemini 3.8 Flashhigh | 56.0 | $0.050 | 7.8 min |
| 10 | Gemini 3.7 Flashhigh | 54.1 | $0.120 | 3.2 min |
| 11 | Kimi K3max | 51.3 | $0.122 | 16 min |
| 12 | Grok 4.6xhigh | 50.5 | $0.180 | 10 min |
| 13 | GPT-6 Lunamax | 50.4 | $0.010 | 6.7 min |
| 14 | GPT-5.6 Lunamax | 48.1 | $0.017 | 5.2 min |
| 15 | Grok 4.5medium | 47.1 | – | – |
| 16 | Gemini 3.6 Flashmedium | 46.9 | – | – |
| 17 | Gemini 3.5 Flash-Litemedium | 25.6 | – | – |