DeepSeek shipped V4 Pro on Wednesday and repriced it on Thursday, splitting its API into peak and off-peak windows. The cheapest line item on the old rate card, the one every multi-turn agent depends on, goes up twelvefold.
Feeding a year of support tickets, every one of them, into a model and asking which complaints actually predicted churn is the kind of analysis that used to mean standing up a data team or writing a check to a research firm. For most of this year it meant neither. DeepSeek's V4 Pro holds a million tokens in a single pass and charges 87 cents per million tokens of output, which is the range where "read everything" stops being a project you scope and becomes something you just run on a Tuesday. At noon Eastern this Sunday, that same output starts costing as much as $3.96, and DeepSeek published the new rate card itself.
Here is what actually changed. V4 Pro left preview and went generally available on Wednesday as the 0813 build, a mixture-of-experts model with a one-million-token context window and a 384,000-token maximum output. A day later the company posted new prices. DeepSeek's own pricing docs now carry both tables, today's rates and the ones that replace them at 16:00 UTC on Sunday, August 16.
The structure is the interesting part, not the size. DeepSeek is not simply raising prices, it is splitting them by time of day, with off-peak rates set at half of peak. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, and everything outside those windows is off-peak. For V4 Pro, output goes from 87 cents per million tokens to $1.98 off-peak and $3.96 at peak. Uncached input goes from $0.435 to $0.66 and $1.32. Cached input, currently $0.003625 per million, goes to $0.022 and $0.044. Reuters put the range across all models and token types at 50 percent to 1,100 percent above current prices, and the published tables bear that out. The smallest increase is uncached V4 Pro input off-peak, up about 52 percent. The largest is cached V4 Pro input at peak, up more than twelvefold.
Notice which column moved the most, because it is not the one anybody is writing headlines about. Cached input is the cheapest thing on the old rate card and the single biggest line item in any agent that runs more than one turn. Every loop resends the system prompt, the tool definitions, and the accumulated conversation, and the cache is what made all that resending nearly free. That is the number going up twelvefold. If you built a multi-turn agent on DeepSeek and modeled its unit economics on cache hits costing effectively nothing, the model is now wrong by an order of magnitude, and the workloads that looked cheapest are the ones getting repriced hardest.
There is a quieter detail here that is worth an hour of somebody's time. Convert those peak windows and they land on 09:00 to 12:00 and 14:00 to 18:00 in Beijing, which is the Chinese working day with lunch carved out of the middle. In Eastern time they fall between 9pm and midnight, and between 2am and 6am. A team working ordinary American business hours never touches peak at all. So the headline increase is not your increase. Yours is the off-peak column: output up about 128 percent, uncached input up about 52 percent. That is still a real jump and you should reprice for it, but it is a different conversation than the top-line number implies. The exception is worth checking today. If any part of your pipeline runs on an overnight batch schedule, that batch now sits squarely inside the expensive window, and moving it is a cron change rather than a project.
Worth noting what else DeepSeek shipped this week. The company released DeepSeek Harness, an open-source agent framework under an MIT license, built so that models, tools, skills, sandboxes, storage, and scheduling are all swappable plugins. Free framework, more expensive fuel. That is not a contradiction, it is a business model, and every other lab in this market is running some version of it.
Two things keep this from being a crisis. The absolute numbers are still low, since even at its most expensive V4 Pro will cost $3.96 per million output tokens, and plenty of teams are paying multiples of that today without flinching. And nobody is trapped, because switching inference providers is closer to a config change than a migration. The sharper problem is what sits underneath. Independent coverage of the GA release notes that DeepSeek's own benchmark claims for the 0813 build have not been replicated by any outside evaluator, and the 0813 weights have not been published, so the escape hatch of self-hosting the exact model you have been calling does not currently exist. Your costs, your throughput, and your ability to verify what you are buying all sit on the other side of somebody else's decision.
Inference has spent two years behaving like Moore's law with better marketing. This week it started behaving like electricity, where the question is not only how much you use but when you use it. Anybody who wrote "and it keeps getting cheaper" into a business case just learned that was an assumption, not a trend.