How we verify prices
Each price on this site comes from the provider’s own pricing page or, for open-weight models sold by other hosts, the cheapest host listed on OpenRouter, and shows the date it was last checked. Automated feeds only tell us when to look again; they never change a price on their own.
Where prices come from
For each model we record the price on the developer’s official pricing page or documentation, a link to that page, and the date we checked it. Open-weight models (gpt-oss, Gemma, Qwen open models, Nemotron and others) are sold by many hosts at different prices; for those we show the cheapest host listed on OpenRouter and say so on the page.
When a provider hasn’t published something (a cache price, a release date, a context window) we leave it blank rather than estimate. A few prices come only from third-party listings; those pages say so.
What “verified” means
The verified date is the last time a person compared our figure with the provider’s page. Every current model was checked between 13 Oct – 11 Nov 2026. Each price row shows its own checked date and links to the page we checked. The dates come from our records, never from the day the site was last built. Previous models that no longer appear on the provider’s main pricing page keep the date they were last checked, and pages that list them show the range of dates.
Catching price changes
Providers change prices with little notice, so we also compare our data with two public price feeds: LiteLLM’s model price list and OpenRouter’s model API. When a feed disagrees with us, or lists a new model from a provider we track, a person checks the provider’s page before anything changes. The feeds are often wrong in their own ways (third-party hosts, stale entries), which is why they can’t overwrite a price.
Pricing details we handle
- Cached input: the per-million price for repeated prompt prefixes, where a provider publishes one.
- Batch: the discount for jobs that return within hours.
- Long prompts: some models charge more above a length threshold (for example Gemini 3.1 Pro above 200K tokens, GPT-5.4 and later above 272K).
- Promotions and peak pricing: temporary prices (such as Gemini 3.8 Flash until 31 December 2026) show their end date; DeepSeek’s peak and off-peak prices are both listed, with peak shown in tables.
- Regions: we show US dollar list prices. Alibaba’s are Singapore-region prices. Taxes are not included. Other currencies are converted at European Central Bank reference rates (currently 25 Sep 2026).
Tiers and scores
Frontier, workhorse and small-and-fast are our grouping by intended use, not a ranking of quality. We don’t publish our own quality estimates.
Capability scores
The capability score is the Epoch Capabilities Index from Epoch AI, an independent research group, which combines a model’s results on many benchmarks into one scale. Higher means more capable on those benchmarks. Each score comes with a range; when two models’ ranges overlap we treat them as level rather than calling a winner. We record a score only when Epoch’s model name matches ours exactly, so newer snapshots and models Epoch hasn’t tested show no score. Benchmarks don’t capture everything, so test the models you’re choosing between on your own work. Epoch AI’s data is used under the CC BY 4.0 licence; where a page shows a score, it credits Epoch AI. Some pages also show the Arena text leaderboard score with its date.
Image, video, speech and free tiers
Image, video and speech APIs are priced per image, per second, per minute or per character, so we show them in the company’s own unit and compare at a stated size or resolution. Where a company bills per token and publishes what an image or a minute works out at, we label that figure as the company’s estimate. Hour-of-speech figures for per-character voices assume a typical speaking rate and are marked as estimates. Free tiers are quoted only as each provider states them, with the date we read the page.
The token price index
The index is the median blended price (3 input : 1 output) of current first-party models in each tier. Its method is on that page.
Local speed estimates
The local LLM calculator estimates generation speed from memory bandwidth, calibrated against published llama.cpp results. Estimates are labelled as estimates; measured results will be labelled with the hardware, backend and settings used.
Provider logos
Logos help you spot each provider’s models at a glance. They are trademarks of their owners and are used only to identify those models; their use doesn’t imply any endorsement. The logo artwork comes from the open-source LobeHub Icons set (MIT licence).
Corrections
We correct errors in the data file, which regenerates every page, and record price changes in the price history.
Using the data
All pricing data is free to reuse under CC BY 4.0: JSON, CSV, changelog. Please credit AI Token Price with a link.