DeepSeek API 4 limits verified 2 Aug 2026
Quota reset forecaster
You hit a wall. When do you get it back?
—
time to reset
Published limits
| Plan | Limit | Window | When exhausted | Evidence |
|---|---|---|---|---|
| All accounts | 500 concurrent connections deepseek-v4-pro concurrency | concurrent, not time-windowed | HTTP 429 | ● Documented |
| DeepSeek API rate-limit docs, fetched 2 Aug 2026. A request occupies a slot from send until the response completes. | ||||
| All accounts | 2,500 concurrent connections deepseek-v4-flash concurrency | concurrent, not time-windowed | HTTP 429 | ● Documented |
| Five times the pro-model allowance. | ||||
| All accounts | Not published Requests per minute / tokens per minute | n/a | n/a | ⨯ Undisclosed |
| Verified absence: DeepSeek documents concurrency only. No RPM or TPM figure appears in the rate-limit reference — unusual among major API providers. | ||||
| Expanded quota | 500 / 2,500 per user_id Per-user-id concurrency | concurrent, not time-windowed | HTTP 429 for that user_id only | ● Documented |
| Accounts granted expanded quota get per-user_id isolation so one user cannot consume the whole account's capacity. Capacity expansion is free on request. | ||||