Don't Trust That LLM API Price Cut Yet: What 47 Real Price Changes Since 2024 Show

Google halved a price in August and scheduled it to double back in January

On August 29 our price monitor saw Gemini 3.6 Flash drop from $1.50 to $0.75 per million input tokens, and from $7.50 to $3.75 for output. A clean 50% cut. Two months later Google's pricing page added a second line: from January 1, 2027, the same model goes back to $1.50 and $7.50.

That is what got me to count. 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 each page, before and after the change. Newer ones come from our own monitor. That gave me 47 changes across OpenAI, Anthropic, Google, DeepSeek, Mistral and xAI.

The short version: prices do mostly go down. But the cuts come early or not at all, some of them come with an end date, and the increases tend to land in the parts of the price list you don't look at. Every change, with links to the archived pages, is in the LLM API price history. Here is what the list says when you count it.

29 cuts, 8 increases, and Anthropic and OpenAI never raised a listed rate

Of the 47 changes, 29 were cuts and 8 were increases. The other 10 were neither: a cached-input price appearing for the first time, an alias such as gpt-4o moving to a newer, cheaper snapshot, a temporary discount starting, a preview model being replaced by its stable version.

Anthropic and OpenAI have no increase on an existing model anywhere in this list. That doesn't mean their newer models are cheap. They usually launch the next model at whatever price they want, and the old one stays where it was. But the row you are paying for today did not go up.

Google is the opposite case. It cut prices 9 times and raised them 5 times, more increases than every other provider combined. If you build on Gemini, check its pricing page more often than the others.

The biggest cuts took 80% off

When a price does drop, it usually drops a lot. Three models had their input price cut by 80% in one go, and three more by 75% or 79%.

Do the sum on a real bill. A team spending $3,000 a month on GPT-5.6 Luna on July 29 was spending about $600 for the same traffic on July 31. Nobody sent them an email about it. If you want to see what a cut like that does to your own numbers, the model switch cost calculator prices your workload on any two models side by side, and it takes about a minute.

If a cut is coming, it comes in the first two months

This is the part I didn't expect. For models launched in 2026, I matched the first cut against the launch date. Every one of them landed within 57 days. The median was 24.

Two caveats. The Gemini 3.6 Flash and GPT-5.6 Sol dates are the day our monitor first saw the new price, because neither provider published an effective date, so the real gap could be a few days shorter. And lots of 2026 models never got a cut at all. This table tells you when a cut tends to arrive, not that one will.

The increases hide in rows you don't read

Eight increases out of 47 doesn't sound like much. Where they landed matters more than how many there were.

• Four of the eight 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. A 100% increase, listed in a single row of the pricing table, with no announcement I could find. Google's page doesn't say whether the old figure was a mistake.
• DeepSeek-V3's launch discount ended on February 8, 2025. Input went from $0.14 to $0.27 (+93%), output from $0.28 to $1.10 (+293%), and cache hits from $0.014 to $0.07 (+400%). The end date was on the pricing page the whole time.
• In September 2025 DeepSeek moved to one price list for deepseek-chat and deepseek-reasoner. For deepseek-chat that meant input going from $0.27 to $0.56 (+107%), and the night-time discount went away at the same time.
• Mistral Large 4 launched on October 6, 2026 with 50% off for two weeks. By October 21 the price had doubled to the standard rate. Ten days later Mistral cut it again, to $0.70 input and $2.10 output, about where the launch discount had been.
• Gemini 1.5 Flash's output price doubled a few days after the model was announced in May 2024. That one never reached a bill, because paid usage only started on May 30.

Then there is Gemini 2.5 Flash, which I didn't count as an increase because the model changed too. When the preview became the stable model in June 2025, input went from $0.15 to $0.30 and the non-thinking output price went from $0.60 to $2.50, a 317% jump. If you had a budget built on the preview price, that was the change that broke it.

Notice the pattern. Half of the increases touched only a cached price. If your app uses prompt caching, a big share of your bill can sit in that one row, and it is the row that moves without a press release. You can check how much of your spend depends on it with the prompt caching savings calculator. It shows the cached and uncached cost next to each other, so a doubled cache price is easy to spot.

Three Gemini prices are already scheduled to double

These aren't in the 47 because they haven't happened yet. Google has published them on its pricing page:

So the 50% cut on Gemini 3.6 Flash was really a four-month discount. Gemini 3.8 Flash launched in September at the discounted price and has never been sold at the full one. If either is in your stack, your January bill doubles with no change to your traffic. We worked through what the Gemini increases do to a monthly bill, including the models that cost the same today and don't go up.

Not every discount ends badly. DeepSeek V4 Pro got a 75% discount the day after it launched in April 2026, and it was supposed to be temporary. On May 31 DeepSeek made it the permanent price. That one went in the customer's favour.

What I'd do before committing to a model

• Budget at the full list price. If the price you see is a launch offer, look for the end date on the pricing page. It is usually there, in small print.
• Read the cached-input and Batch rows, not just input and output. That's where four of the eight increases happened.
• Don't prepay or sign a volume commitment for a model in its first two months. That is when every cut on a model launched in 2026 arrived.
• Re-price your workload when a cut lands. An 80% cut on a model you're not using can make it cheaper than the one you are.
• Subscribe to price changes instead of checking pages by hand. Google alone changed prices 15 times in this list.

For the last two, the model price monitor has a feed with every verified change, and the AI API cost calculator prices a workload on current list prices. Current rates for the models above are on their pricing pages, for example GPT-5.6 Luna, Gemini 3.6 Flash and DeepSeek V4 Pro.

How the 47 changes were counted

40 of the changes come from Internet Archive captures of each provider's own pricing page: the last capture that still shows the old price and the first one that shows the new price. Third-party price lists were only used to find candidates. A change went in only if two official captures showed it. The other 7 come from our monitor, which checks the official pricing pages and records a change once it is confirmed there.

A new model launched at a lower price than the old one doesn't count as a price change here. Neither do model retirements or prices that are announced but not yet in effect. Image, audio-only and embedding prices are left out. Some pricing pages are rendered by JavaScript and have no readable archive, so a few changes are surely missing, mostly from OpenAI's older models in 2025 and from Mistral and xAI. The counts above cover changes up to November 5, 2026.

The archived changes, with the quoted page text for every capture, are a free JSON file in our public AI API pricing dataset. If you find a change that's missing, or one I got wrong, I want to know.

FAQ