DeepSeek is actually raising prices this time, and by a lot: from August 17, both V4-Flash and V4-Pro move to peak/off-peak billing, and the most expensive item jumps 12×. The good news is off-peak costs half of peak, so shifting work off-peak gets some of it back.
This is the full before-and-after price comparison: how much Flash and Pro rise, how peak vs off-peak hours are defined, why cache-hit input gets hit the hardest, and how to save by shifting off-peak.
Before-and-After Price Comparison: Two Tables Tell the Story
One-line takeaway: V4-Flash output at peak is ¥9 — already more than V4-Pro's pre-hike ¥6 — and V4-Pro output at peak hits ¥27. The new prices take effect 00:00 Beijing time on August 17, with peak hours 9:00-12:00 and 14:00-18:00 and off-peak at half the peak price. As of August 15 the old prices still apply (source); all figures are per million tokens:
V4-Flash:
| Item | Before | Off-peak (after) | Peak (after) |
|---|---|---|---|
| Cache-hit input | ¥0.02 | ¥0.05 | ¥0.10 |
| Cache-miss input | ¥1.00 | ¥1.50 | ¥3.00 |
| Output | ¥2.00 | ¥4.50 | ¥9.00 |
V4-Pro:
| Item | Before | Off-peak (after) | Peak (after) |
|---|---|---|---|
| Cache-hit input | ¥0.025 | ¥0.15 | ¥0.30 |
| Cache-miss input | ¥3.00 | ¥4.50 | ¥9.00 |
| Output | ¥6.00 | ¥13.50 | ¥27.00 |
One comparison worth remembering: after the hike, Flash at peak ≈ Pro before the hike — Flash peak output of ¥9 already exceeds Pro's old ¥6, and Flash peak input of ¥3 exactly equals Pro's old rate. So running Flash at peak is no longer automatically cheap.
Increase Breakdown: Cache-Hit Input Gets Hit Hardest
The largest increase is cache-hit input — V4-Pro goes from ¥0.025 to ¥0.30 at peak, exactly 12×. Per-item increases versus the old price:
| Item | Flash off-peak | Flash peak | Pro off-peak | Pro peak |
|---|---|---|---|---|
| Cache-hit input | +150% | +400% | +500% | +1100% |
| Cache-miss input | +50% | +200% | +50% | +200% |
| Output | +125% | +350% | +125% | +350% |
Why is cache-hit input hit hardest? Because it used to be nearly free and was treated as a built-in discount. But coding agents rely on it most — requests that carry the same repository, specs, and history often exceed 80% cache-hit rates. A 12× jump lands right on the most cost-sensitive part of agent workloads, forcing you to treat cache-hit rate as a core metric.
Peak vs Off-Peak: Shifting Off-Peak Halves the Bill
Moving the same job off-peak halves the bill. Peak is only 7 hours (9:00-12:00, 14:00-18:00), while off-peak is 17 hours, and there's no separate weekend rate — weekends are off-peak all day.
Plenty of tasks can shift: log analysis, offline evaluation, batch code review, document extraction, and training-data generation don't need to run during daytime. Scheduling them for nights or weekends automatically halves the price, which beats almost any cache optimization. Real-time agent services can't avoid peak, so they have to lean on cache-hit rates and model downgrades instead.
Choosing Flash vs Pro After the Hike
After the hike, stick to "Flash for daily batch work, Pro for complex reasoning" — plus "shift off-peak whenever you can." Three signals:
- High-frequency batch, completion, and lightweight agent work should stay on Flash; its 2500 concurrency handles volume that Pro's 500 can't.
- Only complex refactors, long-chain reasoning, and knowledge-heavy tasks justify Pro — the ¥27 peak output premium buys deeper reasoning.
- Move anything deferrable off-peak and both models cost half as much.
Making "off-peak + caching" your default beats agonizing over which model to pick.
Alternatives: Third-Party Coding Plans
If the official hike exceeds your budget, a third-party coding plan is the most direct alternative. Volcengine, Tianyi, and Baidu Qianfan all offer fixed bundles with DeepSeek models, billing against plan quota rather than official balance — so they're unaffected by the peak/off-peak hike.
DeepSeek's API is OpenAI- and Anthropic-compatible, so switching in Claude Code, Cursor, or Cline is just a Base URL and API key swap.
Caveats
Before you panic over the peak price, here are three caveats to keep in mind.
- The new prices only take effect 00:00 August 17; August 15-16 still bill at the old rates, so settle your budget now if you want to lock in early.
- The 12× cache-hit increase is a return from promotional pricing to normal commercial rates, not a uniform 12× across everything — check your real cache-hit rate before concluding.
- The official pricing page may change, so confirm against https://api-docs.deepseek.com/quick_start/pricing before committing.