DeepSeek V4 Pro 正式版上线：Agent 能力增强与 API 更新
DeepSeek V4 Pro Released: Stronger Agents and API Updates
概要
DeepSeek 发布 V4 Pro 正式版，APP、网页端与 API 已同步上线。APP 与网页端用户可通过 「专家模式」使用新版模型，API 侧模型名保持不变，调用方无需改动标识即可获得更新。
对已经在使用 V4 系列的团队来说，这次更新的重点有三处：Agent 能力、API 接口形态、 以及计费方式的调整。
Agent 能力增强
正式版重点强化了 Agent 能力，官方称其在生产环境中的性能提升尤为明显。
需要注意公开基准的测试条件：Code Agent 任务的成绩基于 DeepSeek Harness 极简模式，
参数为 max 思考档位、topp=0.95、temperature=1.0。更换框架后结果可能存在偏差，
选型时建议在自己的实际链路上复测，而非直接采信榜单数值。
API 更新
原生支持 Responses API
DeepSeek API 现已原生支持 OpenAI Responses API 格式，并针对 Codex 做了适配。官方文档 提供了一键配置脚本，可直接完成 Codex 的接入。
对存量项目而言，这意味着从 OpenAI 生态迁移的改造量进一步降低——原本需要在接口层做的 格式转换可以省去。
三档思考强度
V4-Pro 与 V4-Flash 的思考模式新增 low / high / max 三档强度，可按任务复杂度选择：
| 档位 | 适用场景 |
|---|---|
low | 简单任务 |
high | 日常 Agent 任务 |
max | 高度复杂的任务场景 |
这一项的实际价值在于成本控制：此前只能在「思考／不思考」之间二选一，现在可以按任务 分层调度，把预算集中到真正需要深度推理的环节。
定价调整
随 V4 全系列正式版上线，API 价格同步调整，并启用峰谷定价——闲时价格为高峰时段的一半。 新价格自北京时间 2026 年 8 月 17 日 0 时起生效。
对批量、离线类任务（数据处理、批量生成、定时任务等）而言，把作业挪到闲时窗口可以直接 降低一半支出，这类调度改造的性价比通常高于模型层面的优化。
本文信息来源于 DeepSeek 官方公告，具体参数与价格以其官方接口文档为准。
Summary
DeepSeek has released the stable version of V4 Pro, available at the same time in the app, on the web and through the API. App and web users reach the new model via “expert mode”; the API model name is unchanged, so callers get the update without touching their identifiers.
For teams already on the V4 series, three things matter here: agent capability, the shape of the API, and how usage is billed.
Stronger agent performance
The stable release focuses on agent capability, which DeepSeek reports as most noticeable in production workloads.
The benchmark conditions are worth noting: Code Agent results were measured with the
DeepSeek Harness minimal mode at the max reasoning tier, topp=0.95 and
temperature=1.0. Results may differ under other frameworks, so re-measure on
your own pipeline rather than taking leaderboard figures at face value.
API updates
Native Responses API support
The DeepSeek API now natively supports the OpenAI Responses API format, with specific adaptation for Codex. The official documentation includes a one-step configuration script for connecting Codex.
For existing projects this lowers the cost of migrating from the OpenAI ecosystem — the format translation that used to sit in the integration layer can be dropped.
Three reasoning-effort tiers
Thinking mode on V4-Pro and V4-Flash now offers low / high / max, selectable by
task complexity:
| Tier | Use case |
|---|---|
low | Simple tasks |
high | Everyday agent work |
max | Highly complex scenarios |
The practical value is cost control. Previously the choice was binary — think or don’t. Now work can be tiered, concentrating budget on the steps that genuinely need deep reasoning.
Pricing changes
Alongside the V4 stable releases, API pricing has been revised to introduce off-peak rates at half the peak price. The new pricing takes effect at 00:00 on 17 August 2026 (Beijing time).
For batch and offline workloads — data processing, bulk generation, scheduled jobs — moving work into the off-peak window halves the bill outright. That kind of scheduling change usually pays off better than optimising at the model layer.
Source: DeepSeek’s official announcement. Refer to their API documentation for exact parameters and pricing.
-- EOF --