Reference prices, updated several times a day: cost per GPU-hour and per million output tokens.
On-demand $/GPU-hr across four standardized GPU configurations, taken cross-provider and rolled up into a single number.
Hyperscalers and neoclouds only, each counted once however many regions it lists. Per SKU, each provider's price is the median of its $/GPU-hr rows (price_per_hour_usd ÷ gpu_count) in a 24-hour window, carried forward for up to 7 days when it misses a scrape. The latest value is the average across providers, and the index is the arithmetic mean of the four SKU averages. Each earlier window is linked to the next by the ratio of average prices across providers priced in both, so a provider joining or leaving does not move the index. Where a provider's price is known only from its published rate card, a change between two reads is spread evenly across that period, because its actual date isn't known; those points are estimates.
On-demand $/GPU-hr for marketplace listings across H200 and B200, kept apart from the GPU-Hour Index because listing prices swing far more than published rates.
Same calculation as the GPU-Hour Index, over marketplace providers. H100 and A100 are left out because SXM listings for them are too sparse to price, and a window is dropped when any constituent has had no listing for 7 days.
Output token price across five open-weights frontier models, averaged into a single $/Mtoken figure.
For each constituent we take the latest OpenRouter-published output price, scale per-token to per-million (×1e6), and average across constituents.
Output token price of the general-purpose flagship from Anthropic, OpenAI, Google and xAI, averaged into a single $/Mtoken figure.
In each 24-hour window, each product line is priced at its newest model on OpenRouter, and the index is the average. A successor replaces its predecessor when it appears, with no adjustment, because moving to a new generation at the same tier is how these prices actually change. Premium tiers above the general-purpose model are left out.
Output token price of the budget model from Anthropic, OpenAI and Google, averaged into a single $/Mtoken figure.
Same calculation as Closed Flagship, over each vendor's budget line. xAI is left out because it sells no current model at this tier.
Neocloud and Hyperscaler lines use the GPU-Hour Index calculation within that provider type. Every chart point covers a trailing 24-hour window, so the latest point is the published index value.
Token lines.Cheapest Seller averages each open model's lowest on-demand seller price, and needs every model priced that day. Blended 3:1 weights input and output prices three to one. Closed Flagship and Closed Efficient price each vendor's newest model at that tier. Closed Premium divides the flagship basket by the open basket. H100 Break-even divides the H100 neocloud price by the open basket, counting output revenue only, at full utilization.
On-demand price per GPU-hour across H100, H200, B200 and A100, averaged across hyperscalers and neoclouds with each provider counted once.
| Name | Last | 1D | 7D | 30D | 6m range |
|---|---|---|---|---|---|
| GPU-Hour Index | $7.19 | +0.5% | +0.7% | +2.6% | $6.30 – $7.19 |