OpenAI API pricing for every GPT model, newest first

A row shows OpenAI's own price where OpenAI sells the model. Where it doesn't anymore, like some older Codex and GPT-4 snapshots, the row shows the host that charges least, which is often Microsoft Azure. Input and output are per million tokens.

The newer GPT models also come in a Pro mode, which thinks much longer before it answers. When a Pro mode bills the same rates per token as its base model, the table leaves it out, since its row would only repeat the same numbers. The thinking still bills as output, so a Pro call can cost several times what a normal one does. Price Pro on real answers, not on the table. Older Pro models with their own rates, like o3 Pro and GPT-5.5 Pro, keep their own rows.

GPT Chat Latest points at the Instant model ChatGPT uses for quick replies, at $5 input and $30 output per million. It follows ChatGPT's updates, so the model behind the name changes over time. Pin a dated model instead if your app needs answers that stay the same.

Caching needs no setup. OpenAI caches any prompt past 1,024 tokens on its own, and on GPT-6.1 Terra a cached token costs $0.05 per million, 95% off the normal input rate. Put the parts that never change first, like the system prompt and the tool list, and the user's message last, so the cache can match the start of each call.

Very long prompts cost extra. On GPT-6.1 Terra, a prompt past 272,000 tokens bills $2 input and $7.50 output per million. GPT-6.1 Terra also has the largest window in the family, 1.05M tokens.

Work that can wait goes through the Batch API, which returns results within 24 hours. On GPT-6.1 Terra that's $0.50 input and $2.50 output per million, 50% off input. Flex processing sits in between: normal API calls that cost less but answer slower and can be turned away when OpenAI is busy. The table shows standard rates only.

The gpt-oss models are open weight under the Apache 2.0 license, so many hosts run them. gpt-oss-120b shows Near AI, at $0.03 input and $0.15 output per million, and you can run it on your own GPUs instead.

Start with GPT-6.1 Terra and run the same prompts through GPT-5 Nano, the lowest cost model OpenAI sells for the example request. If its answers pass your checks, use it for volume. If they don't, step up one tier, not straight to the top. Check the table for a retirement date before you build on an older model, since OpenAI retires snapshots on a schedule.

Use a Pro mode only where a better answer is worth paying for much longer thinking, and measure it per answer. Keep prompt prefixes stable for the cache, and send anything that can wait a day through the Batch API.