DeepSeek-V4-Pro

GA on DeepSeek's own declaration, in its API Change Log for 2026-08-13: 'The GA release of DeepSeek-V4-Pro has been rolled out on the APP, Web, and API. The API calling method remains unchanged - simply set the model name to deepseek-v4-pro to use the latest version.' DEPRECATION WITHDRAWN BY THE PROVIDER. Until 2026-09-12 this row was held at deprecated, with deprecated_on 2026-09-10 and retires_on 2026-09-14, on DeepSeek's announcement that it would retire V4 Pro at 12:00 Beijing Time on 2026-09-14 and route the name to DeepSeek-V4.1-Flash at the Flash price; that announcement has been reversed and no longer exists on either surface that carried it. DeepSeek states: 'In response to user demand, we have decided to continue providing API services for DeepSeek V4 Pro after September 14, 2026, with the billing method remaining unchanged. We will provide further notice should there be any changes.' THE SIGNAL IS THAT STATEMENT, read first-hand from the raw HTML of api-docs.deepseek.com - not an inference from the model still appearing in the rate table. Where it lives changed on 2026-09-19: from 2026-09-12 until then it was carried BOTH as footnote (2) beside the live rate column on the Models & Pricing page and as a paragraph in the Change Log entry for 2026-09-10; the pricing-page footnote has since been removed - September 14 is past, and footnote (2) on that page is now the peak/off-peak schedule - while the Change Log paragraph is still there, re-read first-hand today. The withdrawal therefore still stands on one first-party surface, not two, and V4 Pro is still listed with live rates in the rate table. deprecated_on, retires_on and replacement are therefore null again, because the provider now states no deprecation, no shutdown date and no successor; the 'we will provide further notice should there be any changes' clause is DeepSeek's own and is why this is a withdrawal rather than a guarantee. DeepSeek gives NO DATE for the reversal itself - it edited its 2026-09-10 Change Log entry in place rather than adding a new one - so the only bound on when it published is this catalog's prose baseline, which held the retirement text on 2026-09-11 and the withdrawal text on 2026-09-12. Separately, and as a correction rather than a provider event: this row was carried as preview until 2026-09-10 on our own inference that it belonged to the 'DeepSeek V4 Preview' generation released 2026-04-24, while DeepSeek had in fact declared GA on 2026-08-13 and this catalog did not act on it. The deprecation made that moot for two days; the withdrawal makes it live again, so the row returns to the status the provider declares rather than to the one we had inferred. PRICES ARE UNCHANGED and DeepSeek states the billing method is unchanged: PEAK per 1M is $0.044 cache-hit / $1.32 cache-miss / $3.96 out, with OFF-PEAK exactly half ($0.022 / $0.66 / $1.98). The fields carry the PEAK column by the convention this catalog adopted at the 2026-08-16 peak/off-peak cutover, because DeepSeek presents peak as the rate and off-peak as half of it. The peak window is narrow and got narrower on or just before 2026-09-19: 'Peak hours are 01:00 - 04:00 and 06:00 - 10:00 UTC, Monday through Friday, excluding Chinese public holidays. All other hours are off-peak, including weekends and Chinese public holidays in full.' Off-peak therefore covers every weekend hour, every Chinese public holiday in full, and most weekday hours. Until that revision the footnote read 'Peak hours are 01:00 - 04:00 and 06:00 - 10:00 UTC, Monday through Friday (all other hours are off-peak)', with no holiday carve-out, and no rate moved with the change. Spec, unchanged: largest V4 model (~1.6T total / ~49B active params per its HuggingFace model card and authoritative third-party reports; MIT licence; mixed FP4/FP8), 1M context, 384K max output, three reasoning-effort modes (non-think / think high / think max), JSON output, tool calls, Responses API and Anthropic-format API supported, chat prefix completion, FIM completion in non-thinking mode only, concurrency limit 500, VISION NOT SUPPORTED (the feature table marks it 'Not supported' here and supported on DeepSeek-V4.1-Flash). Model version string: DeepSeek-V4-Pro-0813. Knowledge cutoff never published by DeepSeek, so it stays null. The pricing page is served at the TRAILING-SLASH url (2026-08-11): the extensionless form returns a different document ("Your First API Call") with HTTP 200.

Source: DeepSeek official documentation ↗

Compare DeepSeek-V4-Pro head-to-head

DeepSeek-V4-Pro — questions & answers