Methodology: AI model API prices

AI model API prices

Conditions

What we track. Models in OpenRouter's public model list: the input and output price per 1M tokens, the context window and the maximum output length. New models enter the day they appear in the list; aliases, routers and entries with variable prices are left out; free and batch variants are tracked and marked separately. For flagship models of major vendors we also record the vendor's own API price where its pricing page states it for the named model.

What counts as a value. The price or property OpenRouter lists for the model on the date we read it, in exact decimals; the vendor's price only where its page states one value for the named model in US dollars. Where both exist, we show the difference between them in percent.

What it is not. Listed prices, not a billing test: we do not buy tokens. OpenRouter's price can differ from the vendor's, and top-up fees are not included. Model quality is not rated. What a provider's terms say about your data is read separately, on the terms tracker (see Training on your data below).

Training on your data. The "No training on our data" filter in the model picker comes from the terms tracker at /ai-terms/, not from this price list. For each provider we read one document: Anthropic's commercial terms, OpenAI's API data controls, Google's Gemini API terms, xAI's enterprise terms, DeepSeek's terms of use, Mistral's commercial terms, Alibaba Model Studio's terms and Cohere's terms of use. A model passes the filter only when that document answers "no" to "Do they train on your data?" and the same answer was read on two different dates. An answer read on one date so far is shown on /ai-terms/ with the note "first reading" and is not counted yet. "Depends on plan" does not pass. A provider with no tracked document is shown as not in our data. We report what the document says, not how the provider handles data in practice.

When a value is Unknown. When the list or the vendor page does not publish the value for the model, or the model is no longer listed. A value we have not recorded is shown as Unknown, never as zero.

How often. The list and the vendor pages are read once a day. New and removed models and price changes are listed by date.

AI API Price Watch. The median OpenRouter output price per 1M tokens across paid, non-batch models on the date, with the number of models (N); further series follow flagship models of major vendors, on OpenRouter and at the vendor's own price.

Standard tasks L1, L2 and L3 (Calculated). The calculators, the model comparisons and the cheapest-API tables price the same three tasks. Each figure is the cost of one whole task in US dollars, not a price per 1M tokens.

L1, micro-task (classification or strict JSON extraction): 1 step, 1,500 input tokens, 200 output tokens.

L2, deep analysis over a 50,000-token document: 3 steps; 50,000 input tokens at step 1, 2,500 more at each later step, 500 tokens of tool schema per step, 1,000 output tokens per step.

L3, agentic coding loop: 15 steps; 40,000 input tokens at step 1, 6,500 more at each later step, 2,200 tokens of tool schema per step, 600 output tokens per step.

For each step, cost = input tokens × input price + output tokens × output price, per 1M tokens, at the latest listed price. The plain L1, L2 and L3 figures use no cache. The cached figures assume 75 % (L2) and 85 % (L3) of the step 1 input comes from cache, the whole earlier context is read from cache on later steps, and tasks run 20 an hour so the cache stays warm; they use the provider's listed cache read and write prices, and a price it does not list is charged at the input price. Where the price page or the OpenRouter listing states a higher rate above a token threshold (long context), every step above it is priced at the higher rate. Hidden reasoning tokens are counted as zero unless you enter an estimate in the calculator. A month is 730 hours at 20 tasks an hour.

Workload costs (Calculated, since 11 October 2026). A separate table on /ai-models/workload-cost/ and on model pages prices three everyday workloads on a fixed set of current models: 1,000 chat replies (1,000 input and 300 output tokens each), summarizing 100 documents (4,100 and 400 tokens each) and a day of coding-assistant use (200 requests of 6,000 and 800 tokens). Same arithmetic, no cache, at the vendor's own price and through OpenRouter. Batch and free variants are listed apart and never ranked.

Version. Version 1 of this method, in effect since 29 September 2026; description revised 18 October 2026.

Coverage

On Oct 18, 2026: 466 AI models from 60 providers.

Region

the United States