From Enterprise AI Weekly, recorded 14 August 2026. On Tuesday I said compute gets more expensive from here and that the falling price per token was the wrong number to watch. I got a couple of messages telling me I was being dramatic, which is fair, and I probably was a little. Then three companies repriced in the same week.
DeepSeek raised prices for the first time
I want to be fair to DeepSeek here, because they are a large part of the reason any of us have cheap inference at all. They started the price war, and they are the company everybody cites when they tell you AI is getting cheaper. On Wednesday they announced their first ever price increase: peak and off-peak pricing, live from Sunday, with Chinese business hours costing double the rest of the day. Against the flat rate that preceded it, the peak output price is up more than four times. The wording was that this would allocate resources more reasonably, which is a polite way of saying they are short of compute and have started rationing it with price.
The cache tax
There is a second layer to that announcement I have not seen covered anywhere, and honestly I only found it because I went looking at the cache line specifically. This is a little geeky, so bear with me. The headline model prices went up three to four times. The cached input price went up about twelve.
Cached input is what agents run on. Long system prompts, the same context re-read every turn, retrieval, tool loops, which is precisely the workload everyone is scaling right now, me included. So the increase is concentrated in the one line item that agentic workloads consume most of, and if you built an agent budget on a cheap cache hit, it changed on Sunday. That is the part worth carrying out of this week, and it is the reason I would go back through any 2027 business case that assumes cost per token keeps falling.
And it is not only them
Google shipped Gemini 3.7 Flash on Thursday at $0.75 per million in and $3.75 out, which is a good price. It is also an introductory price that doubles on 1 January, and to their credit they said so in the launch post, so you can plan around a date. Grok 4.6 shipped at $2 in and $6 out, also good, until you cross 200,000 tokens of context, at which point the whole request reprices at double. Not the overflow, the entire request. That is the one that would catch me out, because nobody sets a context length on purpose. It just grows.
What that costs in practice
Take one realistic agent task at 10 million tokens in and 1 million out. The cheapest model on the market runs you about $1.70. The most expensive runs about $150. That is 89 times, for the same job. I ran it twice because I did not believe it the first time. And the cheapest number on that chart expired that weekend, which is worth remembering the next time you read a post explaining to people who do not buy compute that compute is getting cheaper.
An 89x spread on an identical task is a routing decision before it is a procurement decision, and it is the same argument I made about matching the model to the task, only with a wider gap and a deadline attached. It is also why a falling rate card and a rising bill keep coexisting: the menu got cheaper at one end while the workload moved to the other.
The quote I keep coming back to
OpenAI’s enterprise lead told TechCrunch that six months ago every customer conversation was about what the model can do and whether it is good enough. Then he said this:
Our conversations are never about that now. Now the conversations are about we are spending so much. What visibility do you have? What auditability do you have? What token controls do you have?
OpenAI’s enterprise lead, to TechCrunch
That is the company selling you the tokens, describing what its own customers now ask for. They sit on the other side of the invoice from you, and that is what they hear all day. I do not know about you, but I found it more convincing than anything I could have written this week. Customers stopped asking whether it works and started asking what it costs, and in the space of five days three vendors made it cost more.
Visibility, auditability, and token controls is a fair description of what a measurement layer has to do, and it is notable that the list came from a vendor rather than from me. Those three words are also, roughly, the order in which enterprises acquire the capability: you see the spend, then you can explain it, then you can bound it. Most of the organisations I speak to are somewhere in the first stage and budgeting as though they were in the third.
So the check this week is narrow enough to run on Monday. Do you know what share of your token spend is cached input, and would you notice if its price moved under you? Do you know which of your agents sit above a 200,000 token context on a normal day, given that the threshold reprices the whole request rather than the excess? And is anything in your stack watching the vendors’ own pricing pages, given that one of this week’s three increases was published with a date on it months in advance?
None of that requires a project. It requires a record of what you are already spending, broken down far enough to answer a question somebody else set, which is the same reason an agent portfolio needs splitting by agent before anyone can say which parts of it earned their keep.
If you think I have this wrong, tell me, I welcome that all day long. And if you would rather see your own number than argue about mine, that is the work.