跪拜 Guibai
← Back to the summary

DeepSeek V4-Flash Lands with Native Responses API and a Codex Setup Walkthrough

Foreword

On July 31, 2026, DeepSeek released the official version of V4-Flash.

Many communities were abuzz with discussion, and netizens' address for Liang Wenfeng changed from Liangzi back to Liang Sheng—the speed of this face-changing is even faster than flipping a book...

However, V4-Pro still needs to wait a bit longer, with an expected release in early August. Hopefully, it will launch on time this time and not be delayed again.

(After all, the previously announced release time was mid-July... -_-)

Benchmarks and Pricing

Although V4-Pro has not been officially released yet, based on the currently published results, the official version of V4-Flash has shown significant improvement in Agent capabilities and far surpasses the V4-Pro preview version. Here is the official benchmark chart for you to feel:

Benchmark item descriptions in the figure:

Pricing

The official version has the same price as the preview version, but the Agent benchmark scores have significantly improved. It remains a definitive price butcher.

The only blemish is that V4-Flash currently still does not support multimodal inputs like images. But at this price, what more can you ask for?

Responses API Support

A very important change in this update is: the official DeepSeek interface has added native support for the Responses API.

Previously, to connect a model service that only supported Chat Completions to Codex, the common practice was to use the old version of Codex CLI, or to convert the Chat Completions interface to the Responses interface via a proxy.

The former was inferior to the new version in terms of functionality and stability, while the latter added an adaptation layer, which could cause compatibility issues in areas like tool calls and streaming output.

Therefore, for users who want a complete Agent experience in Codex, native support for the Responses API is a very important improvement.

Next, I will teach you step-by-step how to connect V4-Flash to Codex.

Installing Codex

Currently, there are two ways to use Codex:

  1. Codex CLI
  2. The ChatGPT desktop application integrated with Codex

The CLI installation method is very simple. If you already have Node.js and npm installed, you can execute the following in the terminal:

npm i -g @openai/codex

The desktop version's Codex functionality is now integrated into the ChatGPT application. You can go to the official website to download the version suitable for your operating system and complete the installation.

https://chatgpt.com/download

Generating a DeepSeek API Key

https://platform.deepseek.com/usage

Open and log in to the DeepSeek open platform, ensuring your account has sufficient balance. For a first-time trial, you can top up 10 yuan. Actual consumption depends on call frequency, token usage, and cache hit rate. (DeepSeek's cache hit rate is very high)

Enter API Keys in the left menu bar, select Create API Key, enter a name (e.g., Codex), and then complete the creation.

Note: The API Key will only be displayed in full once. Please copy it immediately and save it properly. If the key is lost, you should revoke the old key and create a new one.

The API Key is a sensitive credential. Do not share screenshots or upload it to public repositories. If you suspect it has been leaked, immediately revoke the old key in the console and create a new one.

Installing cc-switch

cc-switch is an open-source configuration management tool that allows quick switching between third-party API providers used by Agent tools like Codex and Claude Code. It has already gained 123k stars on GitHub.

You can download the version suitable for your operating system from the official Releases page.

https://github.com/farion1231/cc-switch/releases

After installing and opening cc-switch, click the OpenAI icon to enter the Codex configuration area, then click the plus sign to create a new provider configuration.

Select DeepSeek as the provider.

Then enter the API Key you just generated in the API Key field.

Set the default model to deepseek-v4-flash and set the upstream format to Responses (Native).

In the model mapping section, just keep one flash.

Scroll to the bottom of the page and set the reasoning intensity (i.e., the model_reasoning_effort field) in the config.toml configuration area:

Three levels are available: low, high, max.

The higher the value, the deeper the model thinks, the higher the answer quality, the longer the time consumed, and the more tokens may be consumed.

After configuration is complete, click Add in the bottom right corner. After returning to the provider list, find the DeepSeek configuration and click Enable.

Testing

If using Codex CLI, execute codex in the terminal. After entering the interactive interface, type /model and press Enter to confirm whether the current model or provider configuration has been switched.

If you are a desktop user, open the application. The model bar in the input box showing Custom indicates that the application has read the third-party configuration. Send a test message and confirm that a reply is generated normally to verify that the interface is successfully connected.

Start your happy Vibe Coding!

Download Materials

It is recommended to prioritize downloading the client from the official website or GitHub Releases to ensure the latest version and trusted source.

But considering that some readers may have access difficulties, I have also organized the versions used in this article into a cloud drive, serving only as a backup download channel.

Quark Cloud Drive link: https://pan.quark.cn/s/3575c7ff95a4

Extraction code: c3Ar

Finally, attached is a quote used by DeepSeek at the end of the V4 preview version's official account article:

"Not tempted by praise, not frightened by slander, follow the path and conduct oneself uprightly."

— Xunzi, "Contra Twelve Philosophers"

If this article is helpful to you, welcome to like, bookmark, and forward it~