the AI bench
VERIFIED AUGUST 2026
All fast takes

FAST TAKE · 2026-08-16 · DEEPSEEK V4 (PEAK / OFF-PEAK BILLING)

DeepSeek starts charging by the clock, and quietly raises its price doing it

At 16:00 UTC on August 16, DeepSeek switched from a single flat rate to peak and off-peak pricing, exactly as it said it would three days earlier. The interesting part is not the mechanism. It is that the new off-peak rate — the cheaper of the two — is more than 50% above the flat rate it replaced.

Verdict: DeepSeek moved to time-of-day pricing on August 16 — and the cheap rate is now higher than the flat rate it replaced


The take

The published card, verbatim from DeepSeek's own API docs: *"Off-peak rates are half of the peak rates. Peak hours are 01:00 - 04:00 and 06:00 - 10:00 UTC (all other hours are off-peak)."* Per 1M tokens, `deepseek-v4-pro` is now **$0.66 in / $1.98 out off-peak** and **$1.32 / $3.96 peak**; `deepseek-v4-flash` is **$0.22 / $0.66** and **$0.44 / $1.32**. Cache-hit input is dramatically cheaper on both — $0.022 and $0.007 off-peak respectively.

Compare that to the flat rate we recorded on August 13, before the switch: pro was $0.435 in / $0.87 out, flash $0.14 / $0.28. So off-peak pro input rises ~52% and off-peak pro output rises ~128%. At peak, output is more than 4.5× the old flat rate. Framed as "we are introducing discounted off-peak pricing," this is a price increase with a discount window attached — and worth naming as such, because the announcement framing invites the opposite reading.

The window placement is the practical detail nobody publishes. Peak is 01:00–04:00 and 06:00–10:00 UTC — two blocks, not one. Map those onto a working day: for a US Eastern developer they fall between 9pm and 6am local, so ordinary US working hours sit **entirely** in off-peak. For a developer in Central Europe, 06:00–10:00 UTC is 08:00–12:00 local, so the whole European morning is peak. The windows line up neatly with Chinese working hours, which is presumably the point. Where you sit determines whether this is a 52% rise or a 128% one.

What we are doing about it: **nothing, for the third sweep running, and now permanently.** DeepSeek still gets no row in the cost calculator. The calculator budgets against a single dollars-per-million figure, and there is no honest way to collapse a two-rate clock into one number without inventing a usage distribution — one that, as above, changes entirely with the reader's timezone. We have held DeepSeek out twice already for a weaker version of this reason. This is no longer a wait-and-see: the rate card is stable and published, and it is still not something a single number can represent. That is recorded in `costCalc.js` so future sweeps stop re-opening it.

Where this fits

Models: DeepSeek V4-Pro · DeepSeek V4-Flash

Hardware: NVIDIA RTX 5090 · NVIDIA RTX 4090

Sources

Next step

Try this in the planner