Statistics · updated daily

AI model statistics, September 23, 2026

Every statistic below is computed directly from the same real data behind the rest of this site - never hand-written - so it updates automatically and never drifts. Free to cite with attribution; link directly to any numbered stat with its anchor.

1. 01

   Anthropic: Claude Opus 5.5 (batch) leads on real quality benchmarks with an Artificial Analysis Intelligence Index of 57.6, among 94 models with published scores, as of September 23, 2026.

   #quality-leaderFull data →
2. 02

   Ox Alpha handles the largest real share of production traffic on OpenRouter, at 27.4% of tracked token volume, averaged over the trailing 80 days.

   #usage-leaderFull data →
3. 03

   The median Intelligence Index across 94 benchmarked models is 23.5 - half of tracked models score higher, half lower.

   #median-quality
4. 04

   inclusionAI: Ling-2.6-flash is the cheapest tracked model at $0.0150 per million blended tokens.

   #cheapestFull data →
5. 05

   OpenAI: o1-pro is the most expensive tracked model at $262.50 per million blended tokens - a 17500x spread from the cheapest.

   #most-expensiveFull data →
6. 06

   The median blended price across tracked models is $0.85 per million tokens.

   #median-price
7. 07

   inclusionAI: Ling 3.0 Flash delivers the best measured quality per dollar, at 654.0 intelligence-index points per dollar.

   #best-valueFull data →
8. 08

   Open-weight models handle 78% of real token volume on OpenRouter, as of September 22, 2026.

   #open-shareFull data →
9. 09

   Anthropic models make up 5% of tracked models by model count.

   #anthropic-shareFull data →
10. 10

    OpenAI models make up 17% of tracked models by model count.

    #openai-shareFull data →
11. 11

    Real production usage is classified across 29distinct task categories, measured directly from OpenRouter's classified traffic.

    #task-categoriesFull data →
12. 12

    DeepSeek: DeepSeek V4 Flash 0423 leads classification with 17% of classified production usage, as of September 23, 2026.

    #classification_taggingFull data →
13. 13

    DeepSeek: DeepSeek V4 Flash 0423 leads data extraction with 16% of classified production usage, as of September 23, 2026.

    #data-extractionFull data →
14. 14

    Z.ai: GLM 5.3 Flash (batch) leads workflow execution with 12% of classified production usage, as of September 23, 2026.

    #agent-workflow_executionFull data →
15. 15

    DeepSeek: DeepSeek V4 Flash 0423 leads roleplay & fiction with 32% of classified production usage, as of September 23, 2026.

    #roleplay_fictionFull data →
16. 16

    DeepSeek: DeepSeek V4 Flash 0423 leads content writing with 31% of classified production usage, as of September 23, 2026.

    #content_writingFull data →
17. 17

    Tencent: Hy-MT2-1.8B leads translation with 19% of classified production usage, as of September 23, 2026.

    #translationFull data →
18. 18

    DeepSeek: DeepSeek V4 Flash 0423 leads data transformation with 24% of classified production usage, as of September 23, 2026.

    #data-transformationFull data →
19. 19

    Z.ai: GLM 5.3 Flash (batch) leads code generation with 19% of classified production usage, as of September 23, 2026.

    #code-general_implFull data →

Real usage vs. real quality

Average usage share (trailing 80 days) plotted against Artificial Analysis quality score, for every model with both. No composite score, just two real numbers side by side - the gap between them is the whole story: Anthropic: Claude Opus 5.5 (batch) tops quality but ranks #339 by usage; Ox Alpha tops usage with a quality score of unpublished.