If you build on an AI API, you have until 30 November before the first of four published price increases lands. That is two weeks from the day this goes up.
Not a rumour, not a leak, not an analyst note. Google prints the new price in the same table cell as the current one. OpenAI prints the expiry date. Anthropic already made its change and nobody wrote it down as a price rise, because on paper it was a cut.
I read all three vendors' own pricing pages on 16 November 2026. Here is what is coming and when.
The four dates
| Date | Vendor | What changes |
|---|---|---|
| 30 November 2026 | OpenAI | GPT-5.6 Sol's promotional pricing ends. New rate not published. |
| 1 December 2026 | Gemini 4 Carbon and Argon double. $2 to $4 in, $10 to $20 out. | |
| 1 January 2027 | Gemini 3.8 and 3.6 Flash double. $0.75 to $1.50 in, $3.75 to $7.50 out. | |
| Already happened | Anthropic | New tokenizer counts about 30 percent more tokens for the same text. |
Three of those four are straightforward. The fourth is the one worth your attention, and I will get to it.
Google tells you exactly what is coming
I want to start with the vendor doing this well, because it is rare enough to be worth saying.
Google's Gemini pricing page does not footnote the increase or bury it in a changelog. It prints both numbers in the same cell. Gemini 4 Carbon's input price reads "$2.00 through November 30, 2026." directly above "$4.00 starting December 1, 2026." Output says $10.00 and then $20.00 on the same dates. Context caching goes from $0.10 to $0.20.
Gemini 4 Argon carries the identical schedule. Then Gemini 3.8 Flash and 3.6 Flash double a month later, on 1 January 2027, from $0.75 and $3.75 to $1.50 and $7.50.
You can budget against that. You know the number, you know the day, and you have time to test whether a cheaper model does your job. Gemini 4 Hydrogen sits at $1.50 and $7.50 with no scheduled change on the page, and it is free on the free tier, which Carbon and Argon are not.
Doubling a price is not generous. Publishing it six weeks out, with the figure attached, is the minimum a vendor owes someone who built a product on their rate card. Google cleared it.
OpenAI gives you the deadline and keeps the number
Now the one that bothers me.
OpenAI's developer pricing page carries this sentence: "GPT-5.6 Sol's promotional pricing is available until November 30, 2026."
That is the whole disclosure. The page does not say what the price becomes on 1 December. It does not even mark which of its published figures is the promotional one. The standard table lists gpt-5.6-sol at $4 input and $20 output per million tokens, and whether that is the promotion or the destination is not stated anywhere I could find.
So you are told to prepare, and not told what for.
The same page does it again with a legal-specialised model, whose promotional pricing is "available at least through December 31, 2026." At least through. No end date, no end price, two pieces of information missing from one sentence.
I am not accusing anyone of anything. Promotional pricing is a normal commercial decision and a vendor is entitled to reprice. But there is a real difference between Google's version of this disclosure and OpenAI's, and the difference costs you the ability to plan.
There is a second thing on that page worth knowing even though it is not a dated change. OpenAI charges a separate long context rate above 272,000 input tokens, and it roughly doubles what you pay. GPT-6 Astra goes from $10 and $50 to $20 and $75. GPT-6.1 Sol goes from $2 and $10 to $4 and $15. If your requests sometimes cross that line and sometimes do not, your effective price is not the number you quoted in your budget.
The price cut that costs you more
Here is the part I think matters most, and the part I have not seen written up anywhere alongside the rest.
Anthropic cut its API prices. These are today's rates, read from their pricing page on 16 November 2026:
| Model | Now | Was |
|---|---|---|
| Claude Fable 5.2 | $8 / $40 | Fable 5.1 at $10 / $50 |
| Claude Opus 5.5 | $4 / $20 | Opus 5 at $5 / $25 |
| Claude Sonnet 5.5 | $2 / $10 | Sonnet 5 at $2 / $10 |
| Claude Haiku 5.5 | from $0.10 / $0.50 | Haiku 4.5 at $1 / $5 |
Every arrow points down. Opus is 20 percent cheaper. Fable is 20 percent cheaper. Haiku looks like a tenfold cut.
And then there is this, on the same page: "Claude 4.7 and later models and Claude Mythos Preview use a newer tokenizer ... This tokenizer produces approximately 30 percent more tokens for the same text."
Tokens are the unit you are billed in. Anthropic cut the price per unit and made the unit smaller.
Run the arithmetic on Opus. You pay 20 percent less per token, so multiply by 0.8. The same document now produces about 30 percent more tokens, so multiply by 1.3. That lands at 1.04. The same work costs roughly 4 percent more than it did on Opus 4.6.
Nobody announced a price increase. Every figure on the page went down. The bill went up.
The caveat, because this cuts both ways
I would be doing exactly what I am complaining about if I left it there, so here is the other half.
This only bites if you are migrating from a model older than Claude 4.7. Opus 4.7, 4.8 and 5 were all already on the new tokenizer at $5 per million input tokens. Moving from any of those to Opus 5.5 at $4 is a clean 20 percent cut with no catch at all.
And one migration gets better rather than worse. Sonnet 4.6 was $3 and $15 on the old tokenizer. Sonnet 5.5 is $2 and $10 on the new one. That is 0.667 multiplied by 1.3, which comes to about 0.87, so roughly 13 percent cheaper for the same text even after the token inflation.
The lesson is not that Anthropic is being sneaky. It is that a rate card stopped being enough information to compare two models, and you now have to know which tokenizer each one uses before the numbers mean anything.
You can see the same change from the other direction in Anthropic's own wording about context windows. A million tokens used to be described as roughly 750,000 words. On the current tokenizer their page puts it at about 555,000. Same window, a third less text in it. I wrote about how much the context numbers vary between a vendor's own pages in what the 1M context claims actually mean, and this is the billing version of the same problem.
The 10 cent model that costs 50 cents
Haiku deserves its own section, because the tenfold cut is real and the headline is still misleading.
Anthropic's pricing page lists Haiku 5.5 twice. Prompts up to 100,000 tokens cost $0.10 per million input and $0.50 output. Prompts over 100,000 tokens cost $0.50 input and $2.50 output.
Five times more, same model, same page.
The page is specific about how the line is drawn: prompt length counts all input tokens including cache reads and cache writes, and each request is priced on its own. So one oversized request is billed at the higher rate, and a workload that mostly sits under the line but occasionally goes over does not average out the way you would hope.
Below 100,000 tokens, Haiku really is about 87 percent cheaper than Haiku 4.5 for the same text, even after the tokenizer adjustment. Above it, the saving drops to roughly 35 percent. Still a cut. Not the one on the front of the page.
This is not the first time I have found a vendor's real price sitting one click below its headline price. I went through seven vendors looking for exactly this in my pricing page versus documentation audit, and the pattern has not changed since.
What genuinely did get cheaper
I have spent most of this post on increases, so let me be straight about the direction of the last two years, because it is strongly downward and the post would be dishonest without it.
Anthropic's own model list still shows Claude Opus 4 and Opus 4.1 at $15 and $75 per million tokens. Today's Opus 5.5 is $4 and $20. That is about 73 percent off per token, and even after allowing for 30 percent more tokens per document it comes to roughly 65 percent less for the same work.
On OpenAI's side, GPT-4o was $2.50 and $10. GPT-6 Luna today is $0.10 and $0.50, with a 1.05 million token context window that GPT-4o never had.
Anthropic also cancelled a planned increase rather than quietly applying it. Their page carries a footnote saying Sonnet 5's $2 and $10 pricing, announced as introductory through 31 August 2026, is now the standard price, and that the scheduled rise to $3 and $15 on 1 September "will not occur." That is the good version of this behaviour: a promotional rate that ended by becoming permanent, with the outcome published.
So intelligence per dollar collapsed through 2026. What changed this autumn is the direction. Three vendors are no longer competing by cutting, and two of them have put dates on it.
What to do before 30 November
Short list, and none of it takes long.
Find out what you actually call. Not what you think you call. Check your logs. I have watched people budget against a model they stopped using in March.
Check whether you cross either cliff. Over 100,000 tokens on Claude Haiku costs five times as much. Over 272,000 tokens on an OpenAI model roughly doubles the rate. Both are per request, so an average will not tell you.
If you are on Gemini 4 Carbon or Argon, your bill doubles on 1 December. Gemini 4 Hydrogen is $1.50 and $7.50 with nothing scheduled, and it is available on the free tier, which the other two are not. Worth testing this week rather than next month.
Screenshot the pricing pages today. Once a price changes, the old figure stops being retrievable, and you will want it when you are reconciling a bill in February. This is tedious advice and I have regretted not taking it.
Count tokens on your real prompts before migrating Claude models. The tokenizer difference depends on what your text looks like. Anthropic's own guidance is to measure rather than assume, and that is the right instinct whatever the vendor.
What I could not check
OpenAI's consumer pricing pages, at openai.com/api/pricing and openai.com/chatgpt/pricing, both returned 403 to an automated request on 16 November 2026. That is the same result I got on 10 November.
A 403 is a server declining to talk to a robot. It says nothing about the product, the company, or the prices, and I have not treated it as evidence of anything. Everything about OpenAI in this post comes from their developer documentation, which is readable. For consumer plan prices I would point you at my ChatGPT Plus review, which I refreshed on 13 November, and I would still check at the checkout screen before paying.
I also did not verify Google's figures beyond the first portion of their pricing page, which is where every number quoted here appears. A later section of that page covers models I have not reproduced.
How I checked this
Every figure in this post came from a vendor's own page, read on 16 November 2026.
Anthropic: the models overview and pricing pages at platform.claude.com. These carry the model lineup, the per-model rates, the Haiku prompt-length split, the tokenizer note and the Sonnet 5 footnote.
OpenAI: the pricing and models pages at developers.openai.com, reached by following a permanent redirect from platform.openai.com. These carry the GPT-6 lineup, the long context threshold of 272,000 tokens and the two promotional notes.
Google: the Gemini pricing page, reached by following a permanent redirect from ai.google.dev to aistudio.google.com. This carries both the current and future prices with their dates.
Where a vendor states something, I have quoted it. Where a vendor is silent, as OpenAI is on what its price becomes in December, I have said so rather than filling the gap from a secondary source. The arithmetic in this post is mine and I have shown it so you can check it.
For the subscription side of the same story, where the pattern through 2026 was flat plans quietly acquiring metered credits, see what AI pricing did in 2026. For what the model names changed to and which ones retired underneath you, see the 2026 model retirements.
One more thing, and it is the only prediction in this post. Price rises published six weeks ahead with the number attached are the best version of this. If the December round passes without much complaint, expect the next round to be announced with less notice, not more.

