Three pricing stories landed this week. They got reported separately. Read together, they point the same way.
What got reported
Google launched Gemini 3.7 Flash at $0.75 per million input tokens and $3.75 output. Roughly half the previous Flash cost. That was the headline.
What the headlines left out
The price is introductory. It runs through December 31, 2026. On January 1 it becomes $1.50 and $7.50.
Google's own pricing page states this plainly, so this is not hidden information. It just did not survive the trip into headlines, which mostly kept the "half price" half of the sentence. If you are sizing a 2027 budget on the current number, you are off by 2x.
The other half of the week
DeepSeek moves V4 Pro to peak/off-peak pricing on August 16 at 16:00 UTC, which is 00:00 Beijing time on Monday, August 17. The full table, per million tokens:
| now | new peak | new off-peak | |
|---|---|---|---|
| cache-hit input | $0.003625 | $0.044 | $0.022 |
| cache-miss input | $0.435 | $1.32 | $0.66 |
| output | $0.87 | $3.96 | $1.98 |
Most coverage led with the output price going up 4.6x. The line that actually moved is cache-hit input: 12.1x.
That is the line agent loops live on. If your system prompt is long and stable, if you are re-sending conversation history every turn, if you are running any kind of ReAct loop, cache hits are most of your input bill. The cheapest line on the sheet just became the fastest growing one.
The timezone nobody converted
DeepSeek publishes its peak windows in UTC: 01:00 to 04:00 and 06:00 to 10:00.
Convert to Beijing time: 09:00 to 12:00 and 14:00 to 18:00. That is the Chinese working day, morning block and afternoon block, with the lunch break carved out of the middle.
Off-peak is exactly half of peak. One thing worth stating clearly, because "half" is easy to misread here: that is half of the new price, not a discount against today. Off-peak cache-hit input is $0.022, still about 6x today's $0.003625. Both tiers cost more than what you pay right now. Scheduling saves you half of the increase, it does not undo it.
The docs give you the UTC numbers. They do not do this conversion, and I have not seen anyone else do it either.
What I would actually change
- On DeepSeek: from 00:00 Beijing time on Monday, August 17, move non-urgent batch work outside 09:00 to 12:00 and 14:00 to 18:00 Beijing time. Same work, half the peak rate.
- On Gemini 3.7 Flash: put $1.50 and $7.50 in the 2027 column, not $0.75 and $3.75.
- On any agent system: re-run the cost model with cache hits at 12x. If cache hits were rounding-error cheap in your spreadsheet, that assumption expires at 00:00 Beijing time on August 17.
The part that does not get repriced
Same week, Qwen shipped Qwen3.8-27B. Dense 27B, Apache 2.0, 262k native context extensible to 1M, and 84.3 on OSWorld-Verified computer use.
I am not going to claim local replaces API. It does not, for most people, most of the time.
But there is one property worth naming while everyone is reading pricing tables: weights you already downloaded do not get repriced on January 1.
Sources: DeepSeek official pricing docs, Qwen3.8-27B model card on Hugging Face, and Google's Gemini API pricing page. The Beijing time conversion is mine, from the UTC windows DeepSeek publishes.
Corrected after publishing: an earlier version said the DeepSeek change starts "Sunday" and that the assumption "expires on Saturday". Both were off. 16:00 UTC on August 16 is 00:00 Beijing time on Monday, August 17. The earlier version also described off-peak as a straight 50% cut without noting it is half of the new price rather than a saving against today's.
Top comments (0)