Jahanzaib

AI model benchmarks

The independent measurements I track, what each one actually tests, and which current models lead it. Every score links to the board that ran it, with the date I checked it. Lab reported numbers are kept out of this page on purpose.

6 benchmarks·157 model scores·last checked Oct 10, 2026

RankingsEvery scored model in one sortable table, with price against intelligence.

Artificial Analysis Intelligence Index

A composite of 10 independent evaluations (agents 30%, coding 20%, scientific reasoning 20%, general 30%), run by Artificial Analysis on each model, version 4.3.2.

by Artificial Analysis · checked Oct 10, 2026

57.6
Claude Opus 5.5
43.4 at $0.20/M
Claude Haiku 5.5
43
  1. 1Claude Opus 5.557.6
  2. 2Claude Sonnet 5.556
  3. 3Claude Fable 5.153.4
  4. 4GPT-6 Astra52.7
  5. 5GPT-6.1 Sol51.8
  6. 6Muse Spark 1.348.1
  7. 7Grok 4.746.5
  8. 8MiMo-V2.6-Pro46.3
  9. 9Qwen3.8-Max45.4
  10. 10GLM-5.344.8
Full board at Artificial Analysis

Arena text rating

Human preference: people compare two anonymous answers and pick the better one. Ratings from the Text Arena overall board, style control on.

by Arena (formerly LMArena) · checked Oct 10, 2026

1507.1 Elo
Claude Opus 5.5
1450.7 Elo at $0.17/M
MiMo-V2.6-Flash
35
  1. 1Claude Opus 5.51507.1 Elo
  2. 2Claude Fable 5.11500.7 Elo
  3. 3Gemini 3.8 Flash1496.5 Elo
  4. 4Muse Spark 1.31494.2 Elo
  5. 5Kimi K31488.1 Elo
  6. 6Gemini 3.1 Pro Preview1487.3 Elo
  7. 7GPT-6.1 Sol1483.7 Elo
  8. 8Qwen3.8-Max1483.3 Elo
  9. 9Gemini 3.6 Flash1482.3 Elo
  10. 10MiMo-V2.6-Pro1480.4 Elo
Full board at Arena (formerly LMArena)

Terminal-Bench 4.0

Agentic coding in a real terminal: can the model, driving an agent harness, finish multi step software and ops tasks. Resolution rate.

by Terminal-Bench · checked Oct 10, 2026

64.9%
Claude Opus 5.5
41.8% at $2.15/M
GLM-5.3
12
  1. 1Claude Opus 5.564.9%
  2. 2Claude Sonnet 5.561.8%
  3. 3GPT-6 Astra58.2%
  4. 4GPT-6.1 Sol58.2%
  5. 5Claude Fable 5.157.9%
  6. 6GLM-5.341.8%
  7. 7Grok 4.737.6%
  8. 8GLM-5.3-Flash35.8%
  9. 9Qwen3.8-Max27%
  10. 10Gemini 3.8 Flash19.1%
Full board at Terminal-Bench

ARC-AGI-2

Abstract reasoning on visual puzzles the model has not seen, semi private set. Hard to memorise, so it tracks reasoning more than knowledge.

by ARC Prize · checked Oct 10, 2026

95%
GPT-6 Astra
59.3% at $0.20/M
GPT-6 Luna
16
  1. 1GPT-6 Astra95%
  2. 2GPT-6.1 Sol94.2%
  3. 3Claude Opus 5.593.3%
  4. 4Claude Fable 5.190%
  5. 5Gemini 3.8 Flash89.2%
  6. 6Gemini 3.1 Pro Preview77.1%
  7. 7DeepSeek-V4.1-Flash72.9%
  8. 8GLM-5.3-Flash65.8%
  9. 9Grok 4.761.4%
  10. 10DeepSeek-V4-Pro61.3%
Full board at ARC Prize

Humanity's Last Exam (Diamond)

Expert level questions across many fields, no tools allowed. The Diamond set is 1,000 questions. Few models are on it yet.

by Scale Labs · checked Oct 10, 2026

60.6%
GPT-6 Astra
55% at $8.00/M
Claude Opus 5.5
9
  1. 1GPT-6 Astra60.6%
  2. 2Claude Opus 5.555%
  3. 3Claude Fable 5.151.3%
  4. 4Gemini 3.8 Flash34.3%
  5. 5Muse Spark 1.325.4%
  6. 6Grok 4.723.4%
  7. 7Kimi K322.2%
  8. 8GLM-5.316.4%
  9. 9DeepSeek-V4-Pro13.4%
Full board at Scale Labs

Output speed

Median output tokens per second measured by Artificial Analysis on the provider API, for the same model setting as its index score.

by Artificial Analysis · checked Oct 10, 2026

364.8 tokens/s
Gemini 3.5 Flash-Lite
n/a42
  1. 1Gemini 3.5 Flash-Lite364.8 tokens/s
  2. 2NVIDIA Nemotron 3.5 Lightning 30B A3B306.7 tokens/s
  3. 3Claude Haiku 5.5240.4 tokens/s
  4. 4gpt-oss-20b226.4 tokens/s
  5. 5Ministral 3 3B221.9 tokens/s
  6. 6DeepSeek-V4.1-Flash217.4 tokens/s
  7. 7Amazon Nova 2 Lite188 tokens/s
  8. 8gpt-oss-120b184.4 tokens/s
  9. 9Muse Spark 1.3175.3 tokens/s
  10. 10Command A+175.3 tokens/s
Full board at Artificial Analysis

How to read these

  • Top score is the best current model on that board. Older and retired models are left out.
  • Best value is the most score per blended dollar (three parts input to one part output, per million tokens), among models within 60% of the leader, so a cheap model that cannot do the task never wins it.
  • No single number picks a model. Shortlist two or three here, then test them on your own tasks. The full directory has prices and context for every model, scored or not.