For two years, every AI project budget was built on the same premise: prices keep falling. In September 2026, one of the market's cheapest providers raised its prices, according to its own founder, without losing customers. For anyone with a project already running, this changes the contract more than it changes the spreadsheet.
What happened
DeepSeek, the Chinese company known for charging a fraction of what American competitors charge, raised the price of using its technology last month. Founder Liang Wenfeng told investors that the increase did not shrink the customer base and that the company's annualized revenue reached 1 billion dollars, up from under 500 million just a few months earlier. The report came from The Information and was confirmed by industry publications.
The size of the increase has circulated as being between 2.3 and 4.5 times, depending on the model. That figure is a report, not our own measurement. What can be verified is the official price table published by the company, and it shows something more interesting than the increase itself.
The new table charges by time of day
DeepSeek currently has two models, and each one costs two different amounts depending on the time of day. Prices are per million tokens — the unit used to measure the text that goes in and out of the system. One million tokens is roughly 400,000 to 700,000 words, depending on the language.
- Fast model: output at $0.60 off-peak and $1.20 at peak. Input at $0.15 and $0.30.
- Advanced model: output at $1.98 off-peak and $3.96 at peak. Input at $0.66 and $1.32.
- Peak hours: 01:00–04:00 and 06:00–10:00 GMT, Monday through Friday. Everything else, including weekends, is half price.
Converted to Brazil time, peak hours fall between 10 PM and 1 AM and between 3 AM and 7 AM. In other words, most of the Brazilian business day falls in the cheap window.
There's another detail worth more than the time-of-day discount. When the same input text repeats — the same manual, the same catalog, the same instruction at the start of every request — the company charges roughly fifty times less for it. This is called caching. A project that repeats the same content on every call and doesn't use this is paying fifty times more than it needs to on that portion.
Does this contradict what we wrote in July?
Partly, yes, and it's worth addressing head-on. On July 30 we published that AI prices fell 80% in three weeks, when OpenAI cut the price of the mid-tier models in its lineup. In that piece we said the decline was likely to continue.
These are two different calculations, and now it's easier to separate them:
- What held up: the top tier doesn't get cheaper. The most capable model kept costing the same in July, and it's still expensive now. Frontier intelligence hasn't become a commodity.
- What fell apart: the idea that baseline prices only move down. They move in both directions, depending on demand and available compute capacity.
The practical lesson isn't choosing between the two stories. It's to stop treating AI pricing as a fixed cost and start treating it as a variable input, just like energy, freight, or the exchange rate.
What changes in your budget
- Put a price-adjustment clause in the contract. If you hire a company that uses AI under the hood, ask what happens if its provider raises prices mid-project. Who absorbs it, and with how many days' notice.
- Measure cost per task, not per month. Knowing that customer support costs a certain number of cents per conversation lets you redo the math in minutes when the price changes. Knowing only the monthly total doesn't.
- Separate what can wait. Classifying documents, summarizing the day's conversations, generating reports: none of this needs to happen live. Running it in batches during off-peak hours can cut that bill in half.
- Use caching for anything that repeats. Fixed instructions, catalogs, price tables: if they're sent with every call, they need to be counted as repeated content, not new text.
- Don't lock the system into a single provider. Switching models should be a configuration change, not a project rewrite. That flexibility is what gives you negotiating power when someone raises prices.
Why they were able to raise prices
The simplest explanation, and the one the industry itself gives, is that there isn't enough compute capacity to go around. DeepSeek allocates more than 70% of its computing power to training new models, and less than 30% to serving users. With demand outstripping supply, prices rise without scaring off customers. The company is also preparing a multibillion-dollar funding round and a stock listing in China, which tends to come with a demonstration of revenue.
For the company doing the buying, the consequence is the same as with any scarce input: whoever depends on a single provider pays whatever price that provider decides. And when the task is hard, it's worth remembering that sometimes it pays off to spend far more on an AI — the math that matters is the cost of error, not the cost of tokens.
October 8 update: prices fell again, but only at the bottom
Twelve days after this article, Anthropic launched Claude Haiku 5.5, which costs on average 75% less to run than the previous model in the same line: US$ 0.10 per million input tokens and US$ 0.50 output for requests up to 100,000 tokens, the same price as OpenAI's GPT-6 Luna. Does this contradict the headline above?
Partly. What DeepSeek's increase showed still holds: AI prices can rise when demand outpaces capacity, and anyone relying on a single provider is exposed. What Haiku 5.5 shows is that at the bottom of the price table, competition is still fierce, as it was in July with the 80% drop in GPT-5.6 Luna. The market does not have one price that only rises or only falls: it has small models getting ever cheaper for high-volume tasks, and contested capacity for everything else.
For your budget, the advice above becomes even clearer: measure cost per task and keep the freedom to switch models. Those who did can take advantage of a 75% drop the day after launch. We ran that math for automated customer service in how much an AI chatbot costs: a thousand conversations come to about US$ 2.25 in AI.
Sources for this update: Anthropic's announcement on the @ClaudeDevs account and VentureBeat coverage (10/7/2026).
How we work
In AI implementation, cost per task is measured in the first week of the project, before any automation goes into production. That number is what lets you switch models, negotiate, or turn off a feature when market prices change. To start the math from scratch, the article on how much it costs to have AI in your business has the current ranges.
Sources
DeepSeek's official price table, checked on September 26, 2026, with per-million-token prices and the peak-hour window. Annualized revenue and the size of the increase: the founder's statement to investors, published by The Information and summarized by the Daily Journal newsletter on September 24, 2026.
Frequently asked questions
Is the price of artificial intelligence falling or rising?
Both. Prices for simpler models dropped sharply in 2026, but in September DeepSeek raised its rates and, according to the founder, didn't lose customers. The most capable model from each provider has barely come down at any point. In practice, AI pricing behaves like a variable input, not a fixed cost.
How much does it cost to use the DeepSeek API today?
Per the official pricing table, the fast model costs $0.60 per million output tokens off-peak and $1.20 at peak. The advanced model costs $1.98 and $3.96 under the same conditions. Input is much cheaper than output, and cached repeated text costs roughly fifty times less.
What counts as peak hours for AI billing?
It's the window in which the provider charges double because of demand. For DeepSeek, peak runs from 01:00 to 04:00 and 06:00 to 10:00 GMT, Monday through Friday. Outside that window, including weekends, prices drop by half. In Brazil's time zone, most of the business day falls in the cheaper band.
How can a company cut AI costs without losing quality?
Separate out what can run in batches outside the expensive hours, use caching for content that repeats across every call, measure cost per task rather than total monthly spend, and match a simple model to a simple task. Save the expensive model for work where a serious error would be costly.
What should go in the contract to guard against a price hike?
Spell out who absorbs a mid-project price increase from the AI provider, how many days' notice a change requires, and what it would cost to switch to another model. Without that, the increase just shows up on the invoice, with no warning and no fallback.
Is it worth relying on a single AI provider?
No. Switching models should be a configuration change, not a system rewrite. That flexibility is exactly what gives you leverage when a provider raises prices or changes the terms of use.