DeepSeek Is No Longer the Cheapest AI. Here Is What Changed

The price list finally arrived

A week ago we reported that DeepSeek had emailed its API customers to warn of a "significant price increase" — without saying how much, or when. Now we know. When DeepSeek officially launched V4 Pro on 13 August, it also published the new rate card, and it replaces the flat per-token fee with something more familiar from your electricity bill: peak and off-peak pricing.

From 16 August at 16:00 UTC, every DeepSeek model carries two rates, and off-peak is exactly half the peak price. The catch is in the clock. Peak hours run 01:00–04:00 and 06:00–10:00 UTC — in Central European Time that is 02:00–05:00 and 07:00–11:00 CET (03:00–06:00 and 08:00–12:00 CEST). In other words, a good part of the ordinary European working morning falls inside the peak window. Schedule your batch jobs after midday and you pay half.

What it actually costs now

The flagship V4 Pro moves from $0.435 per million input tokens and $0.87 per million output tokens to $0.66 / $1.98 off-peak and $1.32 / $3.96 at peak. On output that is roughly 2.3× more off-peak and 4.5× more at peak. The smaller V4 Flash rises from $0.14 / $0.28 to $0.22 / $0.66 off-peak and $0.44 / $1.32 at peak. The full table is in the official pricing page.

In euros, using the European Central Bank reference rate of 1.1534 dollars per euro (13 August), V4 Pro off-peak works out to about €0.57 in / €1.72 out per million tokens, and roughly €1.14 / €3.43 at peak. V4 Flash lands at about €0.19 in / €0.57 out off-peak. These are indicative figures without VAT or platform fees, but they give an honest sense of the scale.

The cheapest-model crown just moved

Here is the part worth paying attention to. OpenAI cut its mid-tier GPT-5.6 Luna by 80% on 30 July, down to $0.20 per million input tokens and $1.20 per million output. That single move reshuffles the whole "who is cheapest" question.

Compared with DeepSeek V4 Pro, Luna is now cheaper on both sides of the bill — $0.20 against $0.66 on input, and $1.20 against $1.98 on output (both off-peak). DeepSeek's flagship is no longer the bargain it was. But the story is not one-sided: the smaller V4 Flash still undercuts Luna on output, at $0.66 against $1.20 off-peak, and on input the two are essentially tied ($0.22 against $0.20).

So the neat shortcut — "just use DeepSeek, it is cheapest" — now has a three-way answer depending on your workload. Heavy output on a tight budget? V4 Flash. Need the stronger model for complex reasoning? Luna may now cost less than V4 Pro. Want V4 Pro's agent features specifically? Run it off-peak and you will still pay far less than for a frontier model from Anthropic or OpenAI's Sol tier.

What this means for you

For European users there is a second layer to this, beyond price. DeepSeek is a Chinese lab and its API processes data on servers in China, which has always been a GDPR consideration for European businesses handling customer or employee data. When DeepSeek was dramatically cheaper, that gap was easier to justify. Now that an American model such as Luna lands in the same price band — with OpenAI's EU data-residency options on the table — the price argument for routing data outside the EU gets noticeably thinner.

None of this makes DeepSeek a bad choice. Its open-source DeepSeek Harness (MIT licence, developer preview) and the new flexible "reasoning effort" setting are genuinely useful. But the era when one model was both the cheapest and the obvious default is over. From this weekend, choosing an AI model is a budgeting exercise again — and that, for everyone who builds on a shoestring, is the real headline.