How to pay less for AI APIs

Batch what doesn't need an instant answer

OpenAI, Anthropic and Google all price batch requests at half the standard rate, input and output. Results come back asynchronously: within 24 hours at OpenAI, most Anthropic batches in under an hour.

Sources:OpenAIAnthropicGemini

Cache the prompt prefix you repeat

Anthropic bills cache reads at 0.1× the input price, with writes at 1.25× (5-minute cache) or 2× (1-hour). OpenAI discounts cached input by up to 95% depending on the model, automatically. Gemini has a separate cached rate plus an hourly storage fee. Put system prompts, tools and documents first, and keep them identical between calls.

Sources:AnthropicOpenAIGemini

Use OpenAI Flex for slow-but-not-batch work

Flex processing bills tokens at Batch API rates with slower, variable latency, and stacks with caching discounts. It's in beta, on a limited set of models.

Sources:OpenAI

Run DeepSeek off-peak

DeepSeek's off-peak rates are half the peak rates: 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, outside Chinese holidays.

Sources:DeepSeek

Skip the fast lane unless you need it

OpenAI's Fast mode (formerly Priority processing) costs more per token than standard. Use it only where latency earns its price.

Sources:OpenAI

Route easy requests to a smaller model

Small models cost a fraction of frontier ones per token. Classify, extract and summarize with a small model; keep the large one for hard reasoning. Compare live prices in the calculator.

Buy the credits for less

Everything above lowers the token bill. Separately, the price of the credits themselves varies by where you buy them: OpenRouter adds a top-up fee, discount books sell below face value within their depth.