Methodology

Compute Screener, built and run by @RenshuBTC, publishes a daily GPU Price Index for each GPU family and a daily Token Price Index for each AI company. Every value is calculated automatically by the rules below and can be recomputed from the published data.

GPU Price Index

Indices for H100, H200, A100, B200, B300, MI300X and RTX PRO 6000. Hyperscalers such as AWS, Azure, Google Cloud and Oracle are not included. Every price is converted to its GPU's benchmark contract:

1. Collect. Every 2 to 3 hours, the index reads posted on-demand prices from each provider's API or pricing page: rate cards, every ask on the Vast.ai and Lium marketplaces, and Akash's average ask. Other smaller clouds are read from Shadeform's public catalog, which also shows their capacity. Each API, page or catalog counts as one source, and no cloud is read through two for the same GPU. Euro prices are converted at the European Central Bank's rate of the day.
2. Filter. Where a source shows capacity, a provider counts for a GPU only if one of its offers had capacity at some collection that day; its posted prices then all count, so an offer selling out does not read as a price change. Asks more than 2× from their marketplace's median and prices more than 2.5× from the GPU's historical reference are left out. A marketplace counts for a GPU only with at least 3 usable asks from at least 2 hosts.
3. Check. A provider's price that moves more than 35% from its 14-day median is held out until the new level has lasted 3 days, unless the market moved with it: net of the median move of at least 3 other sources, it moved less than 35%. A move shared by most of the clouds read through one catalog, and not by other sources, is treated as that catalog's error and stays held.
4. Standardize. Each price is divided by factors for how it differs from the benchmark: form factor and memory, GPUs per rental, region and CPU. Factors are estimated weekly over 28 days from providers that price both configurations on the same day, and need comparisons from at least 2 providers, or 5 matched groups for region and CPU. Without them, a price with another form factor or memory is left out and other differences are not adjusted.
5. Average. Each provider's prices are averaged, so a provider counts once however many listings it has. A provider with no price on a day is carried for up to 7 days, moved by the other providers' average change.
6. Chain. The index moves by the geometric mean of the price changes of providers priced both on the day and on the last published day:index today = index last × geometric mean of (provider average today ÷ provider average last)Both averages use today's standardization factors, so a recalibration does not move the index, and providers joining or leaving do not move it either.

Publication. A value needs at least 5 matched providers, 3 of them priced that day, so no provider weighs more than 20%. A new provider counts after 14 days of collection; the 15 added by version 4.1 count at once, from their second day of prices. If no value was published in the previous 7 days, the index restarts at the cross-section average: the mean of all provider averages with the most extreme ones clipped.

Also published each day: the cross-section average and the quartiles of the provider averages, and monthly averages of the index in /data/index.json.

History. Live values start on 2026-10-23. Before that, H100 and A100 have daily history from 2025-11-07, rebuilt with this methodology from the posted prices in SkyPilot's open price catalog, 5 to 8 providers a day; only these two have 90 days of it without a break. Earlier, and for H200, B200, B300 and RTX PRO 6000, history is weekly: GetDeploying's on-demand median across providers, in weeks with at least 5. Each earlier series is scaled to meet the later one, so a change of source does not move the index.

Token Price Index

One index for each of Anthropic, OpenAI, Google, xAI, Meta, Moonshot AI, Alibaba, DeepSeek, Z.ai and Xiaomi, the top 10 companies in the Artificial Analysis Intelligence Index in October 2026, with DeepSeek always included. A day's value is the average price of the company's paid models, weighted by their tokens that day:index = Σ (tokens × price) ÷ Σ tokens

• Tokens are each model's daily prompt and completion tokens in OpenRouter's daily rankings of the top 50 models, used only to weight a company's models against each other.
• Prices are the company's own posted prices at base rates, without cache, batch, priority or long-context pricing, blended as (3 × input + output) ÷ 4. Since 2026-10-22 they are read from the company's own endpoint on OpenRouter shortly after each UTC day ends.
• Models that count are paid text models. Free variants, batch tiers, embedding and image models, and OpenAI's and Google's open-weight releases, gpt-oss and Gemma, are left out. A model with published weights counts only while the company itself serves it, since other hosts sell it too.
• Coverage. A value is published only when models with a known price carry at least 80% of the tokens that count.

Real bills are usually lower than this list-price blend, because cached input costs a fraction of the base rate. The 3:1 blend is fixed because public data does not show each model's mix.

History. Daily from 2025-01-01, at each day's price from the dated history at llm-prices.com by Simon Willison, or a retired preview's announced price; a model with neither is priced at today's price, assuming it never changed. Where a company's own records date when it served each version, as DeepSeek's change log, Xiaomi's update log, Moonshot AI's model list and an Alibaba notice do, those dates decide which versions count, at the prices the company announced or else llm-prices.com's. A version served free does not count. Otherwise an open-weight model the company serves today counts from its release, and a day is left empty when versions of unknown service or price carry more than 20% of the tokens. DeepSeek's history uses its off-peak base rates and starts on 2026-08-17, the first full day of its current pricing. Where history meets the live index, on 2026-10-22, both give identical values.

Listing

A GPU or company is listed once it has a price and a month of history, so every row has its 1W and 1M change. DeepSeek is always listed. MI300X is collected but not yet published, because fewer than 5 providers price it.

Trending

The Trending tab lists the GPUs by the strength of their price trend over the last 13 weeks, strongest first, whether the price is rising or falling.

1. Prices. The GPU's index on 13 dates a week apart, ending with the latest value (during the day, the provisional value).
2. Line. A straight line fitted by least squares to the logarithms of those prices. Its slope is the trend's speed, shown per month.
3. Score. The slope divided by its standard error, the t-statistic of the slope: large when the price moved steadily one way, small when it moved little or back and forth.score = slope ÷ standard error of the slope

Strength. Strong: a score of 4 or more, up or down. Moderate: 2 to 4. Weak: under 2, a move noise alone often makes. The arrow shows the direction. A GPU with less than 12 weeks of history has no score and is listed last.

Why this measure. On the index history, from 24 January to 26 September 2026, the GPU with the strongest score kept its direction over the next 4 weeks in 29 of 36 weeks, and following it gained 2.3% on average. The GPU with the largest 1-week change kept its direction in 11 of 26 weeks, and following it lost 1.1%: large weekly moves in GPU rental prices tend to partly reverse, while steady 13-week trends more often continued. One year of six GPUs is a small sample, so this is evidence, not a forecast. Trending is information, not investment advice.

Data

Every value can be recomputed from these public files, kept for five years:

Governance

Inputs. The GPU index uses posted prices and marketplace asks, which anyone can rent at, not executed transactions, which are not public. Indices of executed rentals can differ, usually lower, because large buyers negotiate their rates. Providers weigh equally because their rental volumes are not public.

Control. Values are calculated automatically, never set by judgment, and each run passes automated checks before it is published. A GPU index move above 5% in a day is reviewed, and a review can only confirm a value or correct an error.

Independence. No provider or company pays to be included. A Rent link that can earn a referral fee is marked "Paid link" and never affects the index or the order of providers. Compute Screener and the person who runs it hold no futures, swaps, prediction-market contracts or other instruments that settle on GPU rental or AI token prices, and no direct holdings in the providers or companies on this site.

Changes. A material change to the formula, inputs, filters, benchmark contracts or which providers and models count is announced here at least 14 days before it applies, with its version and date. Comments can be sent to the contact on the Privacy page, and a summary is published with the change. Corrections apply at once and are announced when made. Final values change only to correct an error or when a change of method restates history; each such change is recorded with its date, range and reason in /data/legacy/restatements.json. Latest: on 2026-10-24, version 4.1 added 15 providers, the capacity rule, the marketplace minimum and the market check on large moves, and compares each provider's prices on one set of factors. The owner applied it at once, without the usual 14 days' notice, and its new providers count at once instead of after 14 days of collection. H100 and A100 history was rebuilt with it. Earlier that day, version 4.0 introduced the chained GPU index and daily H100 and A100 history, and the token histories of DeepSeek, Moonshot AI, Xiaomi and Alibaba were corrected from those companies' own records.

Third parties. The calculation runs on GitHub Actions and the site on Cloudflare Pages. If the ECB rate cannot be read, the last rate up to 7 days old is used, and after that euro prices are left out. A source that fails contributes nothing that run; the latest status of each is below.

Licence. The values and data on this site are licensed CC BY 4.0: anyone may reuse them, including commercially, with credit to "Compute Screener (computescreener.com)" (see the terms). Third-party data keeps its own terms: CC BY 4.0 for GetDeploying and OpenRouter, Apache-2.0 for SkyPilot's catalog.

Continuity. An index that cannot be calculated reliably is not published until it can be; its history stays available.

Questions and complaints go to the contact on the Privacy page.

Sources

Status of the latest collection.