AI Coding Cost by Model

Every rate in this table was read on 2026-09-03 on the pricing page of the provider that charges it — never from a third-party comparison. The last column names the page and any condition attached to the number. The rate is a snapshot, not a promise, and this family carries prices only because it hands them straight into an input field. On each model page both prices arrive prefilled and stay editable, and the calculator uses whatever stands in the field. The calculator underneath deliberately holds no price list, for the reason set out in its own file: a page with a rate baked into it goes wrong quietly, keeping a confident number long after the number stopped being true. A prefilled field with a check date next to it is the one shape a price can take on a statically built page without doing that. All eight providers quote per million tokens and say so on the page — Anthropic writes "MTok = Million tokens", Moonshot writes "1M = 1,000,000" — so nothing here needed converting. Some providers quote per 1,000 tokens instead, Amazon Bedrock among them; multiply such a rate by 1,000 before you enter it. Four things these two numbers do not cover. Prompt caching: a cache read typically costs a tenth of the input rate and a cache write 1.25 to 2 times it, so a heavily cached agent runs well under this figure. Batch processing: asynchronous requests are commonly half price. Server-side tools: web search at $10 per 1,000 searches, container and session runtime by the hour, all billed apart from tokens. And tool definitions, which add several hundred to several thousand input tokens to every single request. Same rate does not mean same bill. Claude models from 4.7 on use a newer tokenizer that produces about 30% more tokens for the same text, so a given file costs more at an unchanged per-token price. Gemini 3.6, 3.7 and 3.8 Flash are on a promotional rate that doubles on 2027-01-01, and GLM-5.3-Flash on a launch discount that doubles on 2026-09-09, six days after this check. Gemini Pro and every Grok model step up above 200k tokens of context. DeepSeek's listed rate is the peak-hour rate and halves off-peak. Kimi K3's input rate is the cache-miss rate. Seven models have a row and no page: four are marked retired by Anthropic and priced only on Bedrock and Google Cloud, two are limited availability, and Kimi K2.7 Code has a documentation page that names the model and prints no rate. Guessing that one number is exactly the failure this family exists to avoid. Amazon Bedrock and Cohere are missing for the same reason. Bedrock's tables did not render a Nova figure on the check date, and Cohere prints rates only for its legacy models. No source, no row. The order is by output price, highest first: output is charged at four to five times the input rate on almost every list here, so it is what decides a coding bill.