AI Model Spend Leaderboard

Estimated dollars spent per model, based on real token usage routed through OpenRouter and live per-token pricing. Every figure is an estimate — see methodology for the input/output ratio assumption.

Total est. spend (7d)
$8.09M
Top spender
GPT-6 Astra
Top spender · 7d
$3.94M
Models tracked
11
#ModelProvider Tokens (7d) Blended $/M Est. Spend (7d) Range
1 GPT-6 Astra
general
openai 246.06B $16.00 $3.94M $2.95M – $5.41M
2 MiMo V2.5
general
Xiaomi 2.64T $0.43 $1.15M $911.6K – $1.51M
3 Claude Sonnet 4.6
agentic
Anthropic 260.95B $3.96 $1.03M $939.4K – $1.72M
4 MiniMax M3
agentic
MiniMax 445.37B $1.49 $662.7K $614.6K – $1.02M
5 DeepSeek V4 Flash
chat
DeepSeek 4.66T $0.12 $559.4K $372.9K – $606.0K
6 Gemini 3 Flash Preview
chat
Google 237.86B $1.13 $267.6K $148.7K – $297.3K
7 DeepSeek V4 Pro
agentic
DeepSeek 695.77B $0.32 $219.9K $215.7K – $250.5K
8 Gemini 2.5 Flash
chat
Google 142.26B $0.85 $120.9K $58.3K – $136.6K
9 Gemini 3.1 Flash Lite
chat
Google 161.08B $0.56 $90.6K $50.3K – $100.7K
10 DeepSeek V3.2
general
DeepSeek 133.93B $0.30 $40.3K $38.4K – $43.1K
11 GPT-4o Mini
chat
OpenAI 41.39B $0.26 $10.9K $7.1K – $11.8K

Source: OpenRouter (openrouter.ai/rankings), as of —.

Methodology

How we estimate spend. OpenRouter publishes daily token totals per model (prompt + completion tokens combined). We multiply those totals by each model's published per-token pricing to estimate dollars spent.

The input/output caveat. OpenRouter's dataset does not split tokens into input vs. output, and output tokens typically cost 3–5× more than input tokens. We therefore apply an assumed mix — by default 85% input / 15% output for general traffic, and more input-heavy assumptions (90–92%) for coding and agentic models, reflecting how production agent workloads consume context. Because this is an assumption, every figure is shown with a range: the low bound assumes a 95/5 input-heavy mix, the high bound assumes 70/30.

What this does and doesn't capture. Figures reflect usage routed through OpenRouter only — not direct API traffic to OpenAI, Anthropic, Google, etc. — so treat them as directional, not total-market. Prompt-caching discounts are not reflected, which may overstate spend on heavily cached models. Token counts come from each provider's own tokenizer, so cross-model token comparisons are approximate; dollar math is unaffected.