Loading prices
Figures worked out from the live prices on this site, most surprising first. Each one names its formula and sources and updates as prices change. Quote them freely with a link back.
A year ago the best AI model on sale, Grok 4, scored 87% on GPQA Diamond and cost $6.00 per 1M tokens. Today a model that scores as well costs $0.105.
Bar: the best GPQA Diamond score (Epoch AI) among models with a list price 365 days ago. Then: the cheapest blended price that day among models at or above the bar. Now: the cheapest blended price today among models at or above the same bar. Blended: 3 input to 1 output tokens.
Source: Epoch AI, OpenRouter, LiteLLM price map. As of 2026-10-25. Based on 77 data points.
The cheapest model scoring 70% on GPQA Diamond cost $26.25 per 1M tokens in Dec 2024 and $0.03 on 2026-10-25: ×888 less, or half the price every 10 weeks.
First and latest day of the series "cheapest blended price among models with an Epoch AI GPQA Diamond score of 0.70 or more". Halving time: days between them ÷ log2(first ÷ latest).
Source: Epoch AI, OpenRouter, LiteLLM price map. As of 2026-10-25. Based on 125 data points.
The cheapest H100 hour on sale is a Google Cloud spot VM at $1.09 per GPU, less than any GPU cloud or marketplace (lowest: $1.99, a RunPod community listing). The same machine on demand costs $12.08, 11 times as much.
Lowest spot price per GPU-hour among hyperscaler H100 machines (a3-edgegpu-8g-nolssd, Tel Aviv), against the lowest per-GPU price of every GPU cloud and marketplace offer, community listings included. Spot capacity can be taken back at short notice.
Source: Google Cloud GPU pricing, GPU cloud and marketplace prices. As of 2026-10-25. Based on 5 data points.
GLM 5.2 at fp4 costs $0.034 per 1M input tokens at the cheapest of 8 hosts and $1.40 at the dearest: the same model at the same precision.
For each model, hosts serving the same stated precision (fp4 with fp4, fp8 with fp8) and at least three of them: dearest ÷ cheapest list price per 1M input tokens. The widest gap across models. Promotional prices count as listed.
Source: OpenRouter. Based on 8 data points.
Sending 1 TB from AWS (N. Virginia) to the internet costs $92.16. Hetzner rents a whole 6-core server with 64 GB of memory and unlimited traffic for $48.00 a month.
AWS first-tier internet egress price per GB × 1,024 GB (the first 10 TB of a month). Against the cheapest live dedicated-server offer whose traffic is listed as unlimited, monthly price in US dollars.
Source: AWS data transfer pricing, Hetzner dedicated servers.
At 60% on GPQA Diamond, the cheapest US model costs $0.028 per 1M tokens; the cheapest Chinese model $0.03.
Cheapest blended price per developer country among models with an Epoch AI GPQA Diamond score of 0.60 or more (at least three models per country). Country is the developer’s, from Epoch AI.
Source: Epoch AI, OpenRouter, LiteLLM price map. As of 2026-10-25. Based on 143 data points.
16 AMD EPYC Milan cores with 64 GB cost $178 a month as a bare-metal server from Cherry Servers and $894 on AWS (c6a.8xlarge, 32 vCPU, 64 GB). A three-year AWS reservation paid up front still comes to $342 a month.
Cheapest live bare-metal offer with a third-generation AMD EPYC (Milan, AMD EPYC 7313P) against the cheapest AWS m6a/c6a/r6a instance in N. Virginia (AMD EPYC 7R13, also Milan) with two vCPU per physical core and at least as much memory. On-demand × 730 hours.
Source: Cloudadapt bare-metal prices, AWS EC2 pricing.
On Azure, the same VM costs 2.1 times as much in its dearest region as in its cheapest, at the median across 1,401 VM types.
For each VM type sold in five or more regions: the dearest on-demand Linux price ÷ the cheapest. The median across types, per cloud; the cloud with the widest median is shown.
Source: Azure retail prices. Based on 1,401 data points.
1,529 language models have a public per-token price today, ×2.3 as many as a year ago (656).
Distinct language models with a list price for both input and output tokens on at least one host, on the day.
Source: OpenRouter, LiteLLM price map. As of 2026-10-25. Based on 1,529 data points.
The cheapest model that scores 70% or more on GPQA Diamond costs $0.03 per 1M tokens, blended 3 input to 1 output, on 2026-10-25.
Cheapest current blended price among models with an Epoch AI GPQA Diamond mean score of 0.70 or more.
Source: Epoch AI, OpenRouter, LiteLLM price map. As of 2026-10-25. Based on 125 data points.
The same open model costs ×2.0 as much at the dearest host as at the cheapest, at the median across 174 models, comparing hosts that serve the same stated precision.
For each model with three or more hosts serving the same stated precision (fp8 with fp8, and so on): dearest ÷ cheapest list price per 1M input tokens. Median across models. Promotional prices count as listed.
Source: OpenRouter. Based on 174 data points.
An image from one of the ten best-rated image models costs $0.019 at the median.
Median of the cheapest per-image price (standard size) among the ten highest-rated LMArena text-to-image models.
Source: LiteLLM price map, LMArena. As of 2026-10-25. Based on 2 data points.
The cheapest 10-second 1080p AI video costs $0.80 ($0.08 a second).
Cheapest per-second price at 1080p across video models × 10 seconds.
Source: LiteLLM price map. As of 2026-10-25. Based on 25 data points.
Renting one H100 costs $3,219 a month at the median of every on-demand offer.
Median on-demand price per H100-hour across every cloud and GPU marketplace offer × 730 hours.
Source: Cloudadapt GPU prices. As of 2026-10-25.
Input tokens served from the prompt cache cost 90% less than fresh ones, at the median.
Cached-input price ÷ list input price, per model and host that publish both. One minus the median.
Source: OpenRouter, LiteLLM price map. Based on 2,483 data points.
DeepSeek charges up to 50% less in its off-peak hours (Beijing time).
One minus the off-peak list price over the peak list price, per DeepSeek model and token type, in CNY.
Source: DeepSeek pricing. Based on 6 data points.
Batch requests cost 50% less than live ones, at the median, where a host offers them.
Batch input price ÷ list input price per model and host. One minus the median.
Source: LiteLLM price map. Based on 438 data points.
In Sept 2026, 210 AI list prices went down and 105 went up.
Count of list-price changes in the month across every model, host and unit, by direction.
Source: LiteLLM price map, OpenRouter. As of 2026-09. Based on 315 data points.
At 60% on GPQA Diamond, the cheapest open-weights model costs $0.028 per 1M tokens and the cheapest closed one $0.069.
Cheapest blended price among models with an Epoch AI GPQA Diamond score of 0.60 or more, grouped by whether the weights are published.
Source: Epoch AI, OpenRouter, LiteLLM price map. As of 2026-10-25. Based on 123 data points.