Gemini and Gemma API pricing, newest model first

Rows show Google's paid tier price where Google sells the model, and the host that charges least where it doesn't, which the Gemma rows often show. Input and output are per million tokens.

Google is running a 50% discount on Gemini 4 Carbon right now, which brings it to $2 input and $10 output per million. The table shows list prices, so the deal comes off what you see there. Deals like this come and go, so check the end date on Google's pricing page before you plan around one.

Gemini caches repeated prompt prefixes on its own, and those hits bill at the cached rate. You can also create a cache yourself and pay an hourly storage fee to keep it. On Gemini 4 Hydrogen a cached token costs $0.075 per million, 95% off the normal input rate.

Some models change rate with prompt size. Gemini 3.1 Pro Preview bills a prompt past 200,000 tokens at $4 input and $18 output per million, so a long document can change the bill a lot on that model. Gemini 4 Hydrogen reads up to 1.05M tokens, the largest window in the family.

Gemma is Google's open weight line. Google sells Gemma 4 26B A4B at $0.15 input and $0.60 output per million, and you can also download it and run it on your own hardware. The Gemini Omni Flash rows are different again: they're video generation models, and their token rates don't map to a chat request.

Start with Gemini 4 Hydrogen for agent and coding work, and try Gemini 2.5 Flash Lite on the simple, high volume parts. Keep Gemini 4 Argon or Gemini 4 Carbon for the problems that need real reasoning, and measure the cost per accepted answer, not per token.

Prototype on the free tier if your data allows it, then switch to paid before real users show up. Keep long prefixes stable so caching can catch them, and send anything that can wait through the Batch API.