DeepSeek releases V4-Pro and doubles API rates during peak hours
- DeepSeek-V4-Pro enters general availability on the app, web, and API, with app and web access labeled Expert Mode; the API model names remain unchanged.
- The V4-Pro and V4-Flash models add selectable reasoning effort settings: low for simple tasks, high for daily agent workflows, and max for complex tasks.
- DeepSeek adds native OpenAI Responses API support and says its integration is optimized for Codex with one-click setup.
- DeepSeek introduces peak and off-peak API pricing for the V4 lineup; off-peak rates are 50% below peak rates, starting at 16:00 UTC on August 16, 2026.
Hacker News opinions
Flash peak output pricing jumps to $1.32 per million tokens from $0.28, while the cheapest provider is around $0.16. That is a hefty increase.
Off-peak is half the peak rate, which matters. Most Western workplaces will likely get the cheaper pricing.
Those peak windows, 01:00 to 04:00 and 06:00 to 10:00 UTC, line up with Chinese work hours and night in the US. It makes DeepSeek look mostly domestic-focused.
For Europe and the US, those windows are effectively a 2x increase. I initially misread the table and thought off-peak retained the old prices.
My rough comparison is 2x to 2.5x higher off-peak for Flash and 2x to 4x for Pro, with Pro cache input up 6x. Peak doubles those numbers again.
For Flash off-peak, I get 2.5x for cache hits, 1.57x for cache misses, and 2.36x for output. Peak is 5x, 3.14x, and 4.71x respectively.
The full off-peak table is $0.007, $0.22, and $0.66 per million for Flash cache hit, cache miss, and output. Pro is $0.022, $0.66, and $1.98, then both tiers double at peak.
I wonder whether third-party DeepSeek providers will follow these new rates or try to remain competitive at the old prices.
A price increase alone can push users to other vendors because many people will not calculate the exact multiplier. That should help competitors.
I want the API response to state which service tier applied to each request. I calculate and store per-request costs with my results.
Once peak and off-peak token rates arrive, tokens start looking like electricity or old long-distance minutes. I expect a commodity race on price.
I find the churn tiring. I adopt a tool, then two weeks later there is a major change in price or recommended usage, so the offers feel unreliable.
Some US companies already discount batch jobs with noncommittal completion times. The motive is the same: GPUs have finite service lives, so providers want them busy around the clock.
A bot that queues nonurgent queries until off-peak would be easy to build. I expect people to schedule work around cheap LLM hours.