TL;DR: DeepSeek just announced that V4 API prices will double during business hours (9:00–12:00 and 14:00–18:00 Beijing time). The official V4 launch is mid-July 2026. Even at peak rates, DeepSeek remains the cheapest frontier AI API by a wide margin.
What Happened?
On June 29, DeepSeek emailed its users a notice titled "DeepSeek V4 Official Release Plan and Billing Adjustment". The headline change: time-based dynamic pricing — the first of its kind among major AI API providers.
During peak hours (Beijing time 9:00–12:00 and 14:00–18:00 on business days), all V4 API prices double. Off-peak, prices stay at their current (already insanely low) rates.
Think of it as Uber surge pricing, but for AI tokens. When everyone's coding at the same time, you pay a premium. Run your batch jobs at 3 AM? You get the discount.
The New Pricing Tables
DeepSeek V4 Pro
| Billing Item | Off-Peak (¥/M tokens) | Peak Hours (¥/M tokens) |
|---|---|---|
| Input (cache hit) | ¥0.025 | ¥0.05 |
| Input (cache miss) | ¥3.00 | ¥6.00 |
| Output | ¥6.00 | ¥12.00 |
DeepSeek V4 Flash
| Billing Item | Off-Peak (¥/M tokens) | Peak Hours (¥/M tokens) |
|---|---|---|
| Input (cache hit) | ¥0.02 | ¥0.04 |
| Input (cache miss) | ¥1.00 | ¥2.00 |
| Output | ¥2.00 | ¥4.00 |
Real-World Cost: How Much Does It Actually Cost?
One developer (青小蛙 at Appinn) shared his real usage numbers: running DeepSeek V4 Flash to auto-update a World Cup calendar — scraping scores, goals, and match data — for an entire month:
- Normal day: ~¥2 (about $0.28)
- Heavy day (bug fixes, re-runs): ¥13 (about $1.82)
- Total for the month: ¥60 (about $8.40)
That's right — an automated pipeline running daily for a month cost less than a lunch in San Francisco.
How DeepSeek Stacks Up Against Competitors
Here's the output pricing for top models in USD per million tokens:
| Model | Output $/M | Context Window |
|---|---|---|
| DeepSeek V4 Flash (off-peak / peak) | $0.28 / $0.56 | 1,000,000 |
| DeepSeek V4 Pro (off-peak / peak) | $0.87 / $1.74 | 1,000,000 |
| Xiaomi MiMo V2.5 | $0.28 | 1,000,000 |
| Xiaomi MiMo V2.5 Pro | $0.87 | 1,000,000 |
| MiniMax M3 | $1.20 | 1,000,000 |
| GLM-5.1 (Zhipu) | $4.40 | 200,000 |
| Kimi K2.6 | ¥27 ≈ $3.78 | 262,000 |
| Qwen3.7 Plus | ¥8 ≈ $1.12 | 1,000,000 |
| GPT-5.5 (OpenAI) | $30.00 | 1,100,000 |
| Claude Opus 4.8 (Anthropic) | $25.00 | 1,000,000 |
The closest competitor in the "dirt cheap with 1M context" tier is Xiaomi MiMo V2.5, matching DeepSeek V4 Flash at $0.28/M output — but without peak pricing. If you're cost-sensitive and don't want to think about time-of-day, MiMo might be the safer bet.
Why Is DeepSeek Doing This?
The answer is simple: capacity management. DeepSeek's servers get hammered during Chinese business hours. Rather than turning away users (or letting quality degrade from overload), they're using price signals to smooth demand.
This is actually smart infrastructure economics. AWS, Google Cloud, and every major cloud provider have been doing this for years with spot instances and reserved capacity. DeepSeek is just the first AI API provider to apply the same logic to language model inference.
Expect others to follow. If DeepSeek proves this model works, Kimi and Qwen will almost certainly introduce their own peak/off-peak tiers.
What You Should Do
- If you're price-sensitive: Schedule heavy workloads (batch processing, training data generation, bulk analysis) outside Beijing business hours (9:00–12:00, 14:00–18:00). Your wallet will thank you.
- If you need always-on availability: Budget for peak pricing, or set up a fallback chain — route peak-hour traffic to MiMo V2.5 or Qwen3.7 Plus, off-peak back to DeepSeek.
- If you're on a free tier or low budget: V4 Flash at off-peak rates is practically free. ¥1/M input (cache miss) = you can process 1 million tokens of input and 1 million tokens of output for about ¥3 (41 cents). That's an entire novel, in, analyzed, and out.
Bottom Line
DeepSeek V4 at double the price is still the cheapest frontier AI API in the world. The surge pricing announcement is a feature, not a bug — it signals that DeepSeek has enough demand to need capacity management, and the pricing is transparent about it.
If you're building on Chinese AI APIs (and you should be — the price gap vs. US providers is absurd), DeepSeek V4 Flash remains the default choice for anything that doesn't require multimodal input. Just run your heavy jobs after dinner.
Sources: DeepSeek user email (June 29, 2026), Appinn report, DevTk.AI pricing directory. Prices checked July 3, 2026.