This isn't a small bump — DeepSeek lifted the whole pricing table a notch. The official pricing page posted a notice on August 15, 2026: starting 2026-08-17 00:00 CST, billing switches to peak/off-peak pricing, with off-peak at half the peak rate (source). Peak hours are 9:00–12:00 and 14:00–18:00 CST — right in the middle of your coding and batch hours.
What the New Prices Look Like
Peak rates start at 3× and output goes straight to 4.5×. Peak/off-peak pricing splits the day into two tiers — peak at full price, off-peak at half. All new prices below are in CNY (¥), matching DeepSeek's official announcement.
Peak hours (9:00–12:00, 14:00–18:00)
| Item | V4-Flash | V4-Pro |
|---|---|---|
| Cache-miss input | ¥3 / 1M tokens | ¥9 / 1M tokens |
| Output | ¥9 / 1M tokens | ¥27 / 1M tokens |
| Cache-hit input | ¥0.10 / 1M tokens | ¥0.30 / 1M tokens |
Off-peak hours (all other times)
| Item | V4-Flash | V4-Pro |
|---|---|---|
| Cache-miss input | ¥1.5 / 1M tokens | ¥4.5 / 1M tokens |
| Output | ¥4.5 / 1M tokens | ¥13.5 / 1M tokens |
| Cache-hit input | ¥0.05 / 1M tokens | ¥0.15 / 1M tokens |
Both models keep the same specs: 1M context, up to 384K output, thinking/non-thinking modes, OpenAI and Anthropic format support. Concurrency is unchanged at 2500 for v4-flash and 500 for v4-pro (source).
How Big the Markup Is
At peak, v4-flash input is 3× and output 4.5×, while v4-pro cache-hit input jumped 12×. Side by side, the change looks like this (all CNY):
| Item | Model | Before | Off-peak | Peak | Peak markup |
|---|---|---|---|---|---|
| Cache-miss input | V4-Flash | ¥1 | ¥1.5 | ¥3 | 3× |
| Output | V4-Flash | ¥2 | ¥4.5 | ¥9 | 4.5× |
| Cache-hit input | V4-Flash | ¥0.02 | ¥0.05 | ¥0.10 | 5× |
| Cache-miss input | V4-Pro | ¥3 | ¥4.5 | ¥9 | 3× |
| Output | V4-Pro | ¥6 | ¥13.5 | ¥27 | 4.5× |
| Cache-hit input | V4-Pro | ¥0.025 | ¥0.15 | ¥0.30 | 12× |
The ugliest line is v4-pro's cache-hit input — up from ¥0.025 to ¥0.30 at peak, a 12× jump. Workloads that used to live on cache discounts need a fresh cost model.
Top-up, Billing, Quota and Rate Limits Are Unchanged
Top-up and billing logic didn't change: charges = token usage × unit price, auto-deducted from balance, gift balance spent first, no monthly plan. DeepSeek API has no fixed subscription quota — your account balance is your usage cap, and token usage is visible in real time in the console. Top-up and API key creation both happen on platform.deepseek.com.
| Item | Rule |
|---|---|
| Official entry | platform.deepseek.com console |
| Top-up | Top up balance in the console, as needed |
| Billing | token usage × unit price, gift balance first |
| Quota | No fixed subscription quota, balance is the usage cap |
| Rate limit | v4-flash 2500, v4-pro 500; HTTP 429 when exceeded |
| API key | Create in console; OpenAI / Anthropic formats |
On deals: there's no annual discount or free developer quota — the only real "discount" is cache hits. Repeated system prompts and long chat histories hit the cache and cut input cost by dozens of times; it went up this round, but it's still the biggest cost lever. To connect Claude Code, Cursor or Cline, pick OpenAI- or Anthropic-compatible format and fill in the Base URL and key. Need more concurrency? File a capacity ticket — it costs nothing extra. When balance hits zero, the API stops, so top up as you go instead of hoarding.
How to Control Costs After the Hike
Don't fight peak hours — shift every schedulable and cacheable workload to off-peak. Three concrete levers after this hike:
- Run batch jobs off-peak: off-peak is half price, so log analysis, batch completion and offline tasks belong in evenings or non-peak afternoons.
- Max out cache hit rate: repeated system prompts, multi-turn chats and RAG contexts are your highest-ROI caching targets; cache-hit input is still dozens of times cheaper than cache-miss.
- Use third-party coding plans: DeepSeek has no official subscription, but Volcengine, CtCloud and Baidu Qianfan run DeepSeek models on fixed plans with independent pricing — insulated from this peak/off-peak hike.
Is It Still Worth It?
It depends on your volume and timing — heavy users will probably stay, but you have to crush peak-hour calls. This hike shifts the cost of "working-hours" compute onto pay-as-you-go users. If you call heavily during the day, costs climb fast; if you can move work off-peak and get caching right, the off-peak ¥1.5/¥4.5 input/output is still reasonable.
A few concrete moves:
- Measure your peak vs off-peak volume share first — don't fixate on unit price alone.
- Route peak-hour work to cache hits and the lighter Flash model where possible.
- If budget is tight, switch to a fixed plan like Volcengine to lock a monthly cost.
- Keep watching https://api-docs.deepseek.com/quick_start/pricing — peak windows and prices can change.