DeepSeek confirmed to API users today that DeepSeek V4’s official version launches in mid-July, and it comes with a pricing change that’s worth understanding if you’re building on their API.
The headline is a peak-valley pricing model. Your API costs double during certain hours of the day. Outside those windows, you pay the current rates.
Here is what changes, when, and what the smart play looks like.
The New Pricing Structure
DeepSeek is not raising baseline prices. The regular (off-peak) rates match the current V4 pricing on the API docs. During peak hours, every rate doubles.
For deepseek-v4-pro:
- 1M input tokens (cache hit): $0.003625 regular, $0.00725 peak
- 1M input tokens (cache miss): $0.435 regular, $0.87 peak
- 1M output tokens: $0.87 regular, $1.74 peak
For deepseek-v4-flash:
- 1M input tokens (cache hit): $0.0028 regular, $0.0056 peak
- 1M input tokens (cache miss): $0.14 regular, $0.28 peak
- 1M output tokens: $0.28 regular, $0.56 peak
If you are running V4-Pro at scale, a peak-hour inference run that costs you $87 per 100M output tokens today would cost $174 during peak windows.
When Peak Hours Hit
The peak windows are defined in UTC, so factor in your timezone.
UTC peak hours: 1:00 to 4:00 AM and 6:00 to 10:00 AM.
For Beijing time (UTC+8), that translates to: 9:00 AM to 12:00 PM and 2:00 to 6:00 PM.
That second window covers the core working day for much of Asia and bleeds into early US hours. If your production traffic is global, you will likely hit at least one peak window per day.
Why DeepSeek V4 Has Peak-Valley Pricing
DeepSeek told users the adjustment is about resource allocation and service stability. A peak-valley model encourages developers to shift non-urgent work to lower-demand windows, which flattens the load curve on DeepSeek’s inference infrastructure.
The timing makes sense. DeepSeek just released DSpark, a speculative decoding framework that boosts V4-Flash inference speed by as much as 85% on existing hardware. Faster inference plus surging demand (DeepSeek has been operating at or near capacity since the V4 preview launched) creates the kind of capacity pressure that tiered pricing addresses.
DeepSeek is not alone here. The AI API market has been trending toward usage-based tiered pricing for a while. OpenAI and Anthropic both use concurrency-based limits and premium rate tiers. DeepSeek’s approach is distinct because it ties pricing to time-of-day rather than volume or speed tier.
What This Means for API Users
If you are a DeepSeek API user, the play is straightforward.
Batch your async inference jobs to run during off-peak hours. If your workload tolerates latency, push bulk processing, evaluation runs, and background agents into the off-peak windows and you keep the current rates.
For latency-sensitive production workloads (chatbots, real-time agents, streaming responses) that run 24/7, expect roughly 8 peak hours per day at 2x cost. Depending on your traffic distribution, that works out to a 20 to 30 percent effective price increase if you cannot shift load.
There is also the cache play. DeepSeek already prices cache hits at roughly 1% of cache miss rates. If your system design maximizes cache locality, most of your inference stays at the lowest tier regardless of peak timing.
The Competitive Context
This announcement lands in an already aggressive pricing environment. DeepSeek V4-Pro costs about 1.5% of Anthropic’s Claude Fable 5 for equivalent tasks, according to Deutsche Bank analyst Jim Reid. Even at peak rates, DeepSeek remains dramatically cheaper than US alternatives.
Chinese AI labs have been tightening the pricing screw. Z.ai (formerly Zhipu AI) launched GLM 5.2 this week at roughly one-quarter the per-token cost of Anthropic’s enterprise offerings. Coinbase CEO Brian Armstrong announced his company is switching to Chinese AI models for cost reasons. The trend is clear.
DeepSeek’s peak-valley pricing is a refinement of an already dominant cost position, not a retreat from it. Outside peak windows, DeepSeek stays the cheapest option at scale. During peak windows, it still beats US pricing but the gap narrows.
The Bottom Line
DeepSeek V4 goes official in mid-July. Peak-valley pricing starts with that launch. Current V4 prices become the off-peak rate, peak hours double that.
If you can schedule around peak windows, your costs stay flat. If your workload runs around the clock, expect a modest effective increase. And if you have been waiting to see how the V4 pricing shakes out before committing to DeepSeek at scale, this is your cue: the regular rates are here now, and they are locked in as the off-peak baseline.
Read my DeepSeek V4 vision coverage here for context on what V4 can do. And if you are new to DeepSeek, the full DeepSeek AI review covers the platform from end to end.
DeepSeek said it will send 24-hour email notice before the pricing change takes effect. Watch your inbox.
