Tokenix tracks what one unit of AI intelligence actually costs — quality-adjusted and risk-adjusted — across every major model and provider in the market. One number, updated daily.
The master index is a broad-market average of every model we track — frontier flagships and the commodity long tail weighted 50/50, so the cheaper half of the market counts as much as the frontier. A selection of high-weight constituents is shown below — the full set of 2,763 endpoints is available in the live screener.
| Model | Tier | Input /1M | Output /1M |
|---|---|---|---|
o1-pro openai | S | $150.00 | $600.00 |
o1-pro openai | S | $150.00 | $600.00 |
o1-pro-2025-03-19 openai | S | $150.00 | $600.00 |
embed-multilingual-light-v3.0 cohere | S | $100.00 | $0.0000 |
azure/gpt-4.5-preview azure | S | $75.00 | $150.00 |
twelvelabs.marengo-embed-2-7-v1:0 bedrock | S | $70.00 | $0.0000 |
The AI Compute Price Index is a quality-adjusted, risk-adjusted measure of what one unit of AI intelligence costs across the market — expressed as a single number in dollars per 1M Standard Compute Units.
Like the Consumer Price Index tracks a basket of goods, ACPI tracks a basket of AI compute. When the number falls, AI is getting cheaper. When it rises, something in the market is tightening.
Every model is converted to a common unit — cost per 1M tokens — regardless of modality. Text, voice, image, video and GPU cloud pricing are all normalised to this single scale.
Each model's per-token cost is blended 75% input / 25% output — the standard 3:1 usage assumption — and carries a market-risk factor reflecting the concentration and stability of the provider landscape. The master ACPI is a two-bucket broad-market average: models split into a premium bucket (frontier flagships) and a commodity bucket (the long tail), each averaged on its own, then combined 50/50 so the cheaper half of the market pulls on the index as hard as the frontier — a true market measure rather than a frontier price tag. Tier assignment is a disclosed manual classification, reviewed monthly.
Quality is computed from a HELM-aligned benchmark composite (MMLU, coding, math, reasoning), sourced via standardized leaderboard aggregation and z-score normalised. It powers the intelligence-per-dollar screener (P1). Models without available benchmark data are excluded from that screener but remain in the published ACPI price index.
Benchmarks: HELM-aligned composite (MMLU, coding, math, reasoning) via standardized leaderboard aggregation. Token pricing: provider documentation, verified daily. GPU pricing: Lambda Labs H100 SXM5 market median. Throughput: Hyperstack vLLM benchmark, Llama 3.1 70B.
Market-health signal from provider ARR estimates, API accessibility and funding stability. Full methodology at tokenixindex.com/methodology.