News / Model Watch · DeepSeek
DeepSeek-V4-Pro-0813 Official: V4-Pro 正式版 Lands With Stronger Agents
On 2026-08-13, DeepSeek's official API documentation updated deepseek-v4-pro to the DeepSeek-V4-Pro-0813 snapshot — the V4-Pro 正式版 (GA). The model ID is unchanged, agent capability is enhanced, and DeepSeek moves to peak / off-peak billing from 2026-08-16 (16:00 UTC) — V4-Pro output goes from $0.87 to $1.98 off-peak / $3.96 peak per 1M (peak = 2× off-peak). Here is what API builders actually need to do.
Yes — DeepSeek-V4-Pro-0813 (the V4-Pro 正式版) is live as of 2026-08-13.
DeepSeek's official API docs now read: the deepseek-v4-pro model has been updated to
DeepSeek-V4-Pro-0813; the calling method is unchanged — request
deepseek-v4-pro to get the latest version. Chinese tech media (IT之家, 新浪财经, 东方财富)
confirm the 正式版 went live the evening of August 13, with both API and chat updated, and report the headline improvement as
enhanced agent capability, and comes alongside DeepSeek's 2026-08-06 price-hike notice (“预计涨幅较大”) — DeepSeek moves to peak / off-peak billing from 2026-08-16 (16:00 UTC) — V4-Pro output rises from $0.87 to $1.98 off-peak / $3.96 peak per 1M (peak = 2× off-peak).
Existing OpenAI-compatible clients need no model-ID change.
Background reading: DeepSeek-V4-Pro for API Builders.
What “0813” actually is — the 正式版 (GA)
DeepSeek has shipped the V4 family in stages. Back on 2026-04-24 the official update log noted that only the V4-Flash API had been upgraded at that time (to the Responses API format, with Codex support), while the V4-Pro API and App/Web surfaces were unchanged and the V4-Pro 正式版 would come “as soon as possible.” The DeepSeek-V4-Pro-0813 entry now in the docs is that 正式版 landing.
- Before: V4-Pro tracked as pre-正式版 / “coming soon”; V4-Flash (DeepSeek-V4-Flash-0731) already in public API.
- Now (0813): V4-Pro 正式版 in the official docs; reported live on both API and chat surfaces the evening of Aug 13.
- Model ID: still
deepseek-v4-pro— the version suffix is an internal snapshot label, not a new request name.
What changed (and what did not)
Summarized from DeepSeek official docs + Chinese tech media (IT之家, 新浪财经, 东方财富, 快科技). We do not invent benchmark numbers; unspecified cells mean “not officially disclosed.”
| Dimension | V4-Pro (pre-0813) | V4-Pro-0813 (正式版) |
|---|---|---|
| Status | Pre-正式版; “coming soon” | 正式版 (GA) · API + chat updated |
| Model ID | deepseek-v4-pro | deepseek-v4-pro (unchanged) |
| Headline change | Agentic coding & reasoning baseline | Enhanced agent capability |
| Pricing | $0.435 in / $0.87 out per 1M (flat) | Peak/off-peak from 08-16: out $1.98 off / $3.96 peak |
| Architecture | ~1.6T MoE / ~49B active / 1M context / open weights | Carried forward (not a new arch) |
| DeepSeek-V4-Pro-0813 (per 1M tokens, USD) | Input · cache miss | Input · cache hit | Output |
|---|---|---|---|
| Current (until 2026-08-16 16:00 UTC) | $0.435 | $0.003625 | $0.87 |
| Off-peak (from 2026-08-16) | $0.66 | $0.022 | $1.98 |
| Peak (2× off-peak) | $1.32 | $0.044 | $3.96 |
Source: DeepSeek official pricing page (api-docs.deepseek.com/quick_start/pricing — verified). Peak hours 01:00–04:00 & 06:00–10:00 UTC (Beijing 09:00–12:00 & 14:00–18:00); all other hours off-peak; off-peak = half of peak. New rates take effect 16:00 UTC on 2026-08-16. Context 1M; max output 384K. V4-Flash has its own lower schedule. (Chinese-domestic CNY pricing on platform.deepseek.com differs and is lower.)
What this means for API builders
- No model-ID migration. Keep calling
deepseek-v4-pro; it now resolves to 0813 upstream. Pin with a dated alias only if you need reproducibility across snapshots. - Re-test agent workflows. The advertised improvement is agent capability — re-baseline your tool-use / multi-step prompts rather than assuming parity with the previous snapshot.
- Check the price delta. Output rises from $0.87 to $1.98–$3.96/M under peak/off-peak billing — material at volume. Pull the current rate from DeepSeek's official pricing page and update cost ceilings.
- Via ChinaModelAPI: the relay routes
deepseek-v4-pro, so OpenAI-compatible clients automatically pick up 0813. Confirm the live alias in your dashboard if you maintain an explicit model map. - Keep R1 where it fits. Legacy DeepSeek-R1 remains useful for pure-reasoning workloads; 0813 does not retire it.
Primary sources
- DeepSeek API Docs — News (official: deepseek-v4-pro → DeepSeek-V4-Pro-0813)
- DeepSeek 更新日志 — V4-Flash upgrade & V4-Pro 正式版 context (2026-04-24)
- DeepSeek API Docs — Models & Pricing (official USD, verified: V4-Pro $0.435/$0.87 → off-peak $0.66/$1.98, peak $1.32/$3.96 from 2026-08-16 16:00 UTC)
- 澎湃新闻 — DeepSeek 大涨价（2026-08-06 公告：预计涨幅较大）
- 财联社 — DeepSeek 预告大幅涨价 + 峰谷定价（注：国内 CNY 价，本站用官方 USD 国际价）
- 新浪财经 — DeepSeek API 文档新增 DeepSeek-V4-Pro-0813 (2026-08-13)
- 东方财富 — DeepSeek-V4-Pro-0813 接入 API (2026-08-13)
- IT之家 / 搜狐 — DeepSeek V4 Pro 正式版今日晚间发布，增强 Agent 能力 (2026-08-13)
- 快科技 — DeepSeek-V4-Pro-0813 上线，API 及 Chat 更新 (2026-08-13)
- Our earlier guide: DeepSeek-V4-Pro for API Builders
FAQ
Is 0813 a new architecture?
No. It is the V4-Pro 正式版 refresh — enhanced agent capability plus a move to peak/off-peak pricing — carried on the same ~1.6T MoE / ~49B active / 1M context family.
Will my old calls break?
No. deepseek-v4-pro still works and now points at 0813. No client change required.
What is the new price?
Per DeepSeek's official docs: today V4-Pro is $0.435/M input (cache miss) · $0.87/M output; from 2026-08-16 16:00 UTC it goes peak/off-peak — off-peak $0.66/$1.98, peak $1.32/$3.96 per 1M (peak = 2× off-peak). Peak hours 01:00–04:00 & 06:00–10:00 UTC. (Chinese-domestic CNY pricing is lower.)
V4-Flash vs V4-Pro?
Flash (DeepSeek-V4-Flash-0731) is the cheap high-volume tier with Responses API + Codex support; Pro is the frontier agent/reasoning tier. Both share the 1M context, open-weight family.