DeepSeek V4 Flash API Prices Will Drop on September 10

DeepSeek will reduce API prices for its Flash series at 12:00 Beijing time on September 10, 2026. For DeepSeek V4 Flash, the off-peak price per million tokens will fall to RMB 0.02 for cache-hit input, RMB 1.00 for cache-miss input, and RMB 4.00 for output.

The change partly reverses the V4 pricing increase that took effect in August. Cache-hit input returns to its earlier price, cache-miss input also returns to its earlier price, and output becomes cheaper than today's rate without falling all the way back to the earlier RMB 2.00 level.

The three off-peak price cuts

• Input, cache hit: RMB 0.05 today → RMB 0.02 from September 10, a 60% reduction.
• Input, cache miss: RMB 1.50 today → RMB 1.00 from September 10, about a 33% reduction.
• Output: RMB 4.50 today → RMB 4.00 from September 10, about an 11% reduction.

All figures are off-peak rates per one million tokens. The announcement uses renminbi pricing, so this comparison does not apply a live foreign-exchange conversion that could make the result less precise.

Peak prices will still be twice the off-peak rate

DeepSeek is keeping the same time-of-day structure. Peak hours run from 09:00 to 12:00 and 14:00 to 18:00 Beijing time, Monday through Friday. Calls outside those windows, including weekends, use the off-peak rate.

That makes the new peak prices RMB 0.04 per million cache-hit input tokens, RMB 2.00 per million cache-miss input tokens, and RMB 8.00 per million output tokens. The September 10 change starts at noon in Beijing, exactly when the morning peak window ends.

Cache-heavy agent sessions receive the largest cut

The 60% reduction applies to input tokens that hit DeepSeek's context cache. Long coding-agent sessions repeatedly send system instructions, repository context, tool results, and conversation history, so a high cache-hit share can make this line item more important than the headline uncached-input rate.

The practical saving still depends on workload shape. A request with mostly new context will benefit mainly from the 33% cache-miss reduction, while an output-heavy reasoning task receives the smaller 11% reduction. Token price alone does not guarantee a lower total task cost if the model generates more tokens or needs additional retries.

The public price page still shows the current schedule

At the time of writing on September 8, DeepSeek's public API pricing page still lists the rates in force before the September 10 change. Developers should treat RMB 0.02, RMB 1.00, and RMB 4.00 as scheduled prices until the effective time, then confirm the live rate card before updating budgets or customer-facing cost estimates.

The notice refers to the Flash series. It does not announce a V4 Pro price change, a new permanent V4.1 model ID, or a different API endpoint. Existing integrations can keep using deepseek-v4-flash; the change is a billing update rather than an API migration.

What developers should change

No request-schema change is required. Teams that route batch jobs by cost can refresh their rate tables at the effective time and keep peak-hour multiplication separate from base off-peak prices. Cost dashboards should continue to split cache-hit input, cache-miss input, and output instead of applying one blended token rate.

The clearest before-and-after check is to export those three token counts for a representative week and reprice the same usage under the new schedule. That shows the actual saving for a workload without assuming every token receives the largest advertised reduction.

Sources

• DeepSeek Open Platform pricing notice dated September 8, 2026, reproduced in the translated chart above
• DeepSeek API models and pricing
• DeepSeek V4 Pro release note introducing the current peak and off-peak schedule