What AI intelligence costs
Cuts vs substitution
Same-model price (chained), basket turnover (composition) and the all-model average, Sep 2023 = 100, log scale. Ticks mark weeks with a price change in a lab's own list.
Price spread
Blended $/1M across every model with an Epoch AI capability score: median and 10th–90th percentile band, log scale.
Implied margin
Serving cost = GPU rent ÷ DeepSeek's published throughput per GPU; floor Vast.ai H100 at full load, ceiling AWS on-demand at 40%. Margin = 1 − cost ÷ list price. Method
Open vs closed
Cheapest blended list price among models scoring at least the chosen ECI. Open weights shown twice: the lab's own price and the cheapest third-party host. DeepSeek floor lift = the cheapest price without DeepSeek ÷ the cheapest overall, at lab prices and at host prices (1.0× = DeepSeek is not the floor).
Hosts undercut the labs
Open-weight models sold both by their lab and by third-party hosts (DeepInfra, Together, Fireworks and others). Discount = the lab's own price ÷ the host price, blended 3:1. Bars: models at ECI 145+ with three or more hosts.
Leaving the APIs
$/1M tokens, blended 3:1. Desk boxes run one open model for one user: Mac Studio (Apple lease) with GLM-5.3, RTX 5090 (bought, 3 years) with Qwen 3.8 27B; speeds are estimates. Rented and owned H100s serve many users at 40% utilisation on DeepSeek's published throughput. The dashed line is the cheapest closed model at the same capability. Downloads: Hugging Face, last 30 days. Method
Cost per task
Eight short tasks with checkable answers, each model at its default settings. Cost is what OpenRouter charged, per 1,000 tasks; hover for tokens per task, reasoning share and accuracy. Per-token price and per-task cost can move in opposite directions when models think longer (Zhu 2026, arXiv:2608.29843).
Who pays for the compute
Capital expenditure (cash paid for property and equipment) per calendar quarter. Quarters are derived from year-to-date 10-Q and 10-K figures. Oracle's fiscal quarters are placed in the calendar quarter they end in. Lighter bars: not every company has reported yet.
Who collects
The five payers' combined capex against what the market pays the two kinds of collector. Chips and memory: whole-company revenue of NVIDIA, Broadcom and Micron, which includes other customers, so an upper bound. Cloud rental: AWS, Google Cloud (includes Workspace), Microsoft's server products and cloud services (Azure plus server software), Oracle Cloud Infrastructure and CoreWeave. Fiscal quarters are placed by their midpoint.
Capex, tokens and price
All three indexed to Q1 2025 = 100, log scale. Capex: the five payers combined, complete quarters only. Tokens: routed through OpenRouter, full quarters only. Price: average usage-weighted paid price.
Cost of capability
Cheapest $/1M for any model scoring at least 130, 145 or 160 on Epoch's index (Claude 3.5 Sonnet = 130, GPT-5 = 150). Log scale.
Routed token volume
Tokens routed through OpenRouter per week, stacked. The eight largest labs by total volume have their own colour; "Other labs" is every other named lab; "Long tail" is OpenRouter's own total for models outside each day's top 50. Regions follow the lab's headquarters.
Price events
| date | event | model | provider | old $/1M | new $/1M | change |
|---|
Launches, cuts and increases in the labs' own price lists; daily from 23 Oct 2026, weekly before. Full history on the timeline.