Google cut a Gemini price in half in August. It doubles back in January.

Posted on Nov 5 • Originally published at aicostbudget.com

On August 29 Gemini 3.6 Flash dropped from $1.50 to $0.75 per million input tokens, and from $7.50 to $3.75 for output. Two months later a new line showed up on Google's pricing page: from January 1, 2027, it goes back to $1.50 and $7.50.

That made me curious how often this kind of thing happens, so I counted. I took every LLM API price change I could prove from the providers' own pricing pages since 2024. Older ones come from Internet Archive copies of the page from before and after the change. Newer ones come from a monitor I run that checks the official pages. I ended up with 47 changes across OpenAI, Anthropic, Google, DeepSeek, Mistral and xAI.

Short version: prices mostly go down. But cuts come early or not at all, some of them have an end date, and the increases land in the parts of the price list people skip.

The other 10 were neither, for example a cached-input price added for the first time, or an alias moved to a newer snapshot.

OpenAI and Anthropic never raised a listed rate on an existing model. They just launch the next one at whatever price they like. Google raised prices five times, more than every other provider in the list put together.

Three models had their input price cut by 80% in one step: Codestral (Sep 2024, $1.00 → $0.20), o3 (Jun 2025, $10 → $2) and GPT-5.6 Luna (Jul 2026, $1.00 → $0.20). Gemini 1.5 Flash lost 79%, and DeepSeek V4 Pro and Mistral Large 3 lost 75%.

Here's what that looks like on an invoice. A team spending $3,000 a month on Luna on July 29 was spending about $600 for the same traffic two days later, and nobody emailed them about it.

This was the part I didn't expect. I matched each 2026 model's first cut against its launch date:

Every one landed within 57 days, and the median was 24. Plenty of models never got a cut at all, so this is about when, not whether. It still means the first two months after a launch are when you're most likely to be overpaying. That's a bad time to buy prepaid credits or sign a volume commitment.

Four of the eight increases were the same change on four Gemini models in September 2025. The Batch-tier cached-input price went from half the standard cached price to the full price. That's a 100% increase, in one row of the pricing table, and I couldn't find any announcement.

The rest were mostly launch discounts ending. DeepSeek-V3's output went from $0.28 to $1.10 (+293%) in February 2025, though the end date had been on the page the whole time. Mistral Large 4 launched at 50% off for two weeks, doubled, and then got cut again ten days later.

If you use prompt caching, a big chunk of your bill can sit in that cached row. It's the row that changes without a press release.

The full list, where every change links to the archived pages it came from, is the LLM API price history. The longer writeup, with the scheduled Gemini increases and the method, is here. The raw data is a JSON file in a public GitHub dataset if you'd rather count it yourself.

If you know of a price change I missed, especially an older OpenAI or Mistral one, leave a comment and I'll check the archives.