跪拜 Guibai
← All articles
AI Coding · DeepSeek · Agent

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

By 极客密码 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Native Responses API support turns a cheap, high-performing Chinese model into a drop-in backend for OpenAI’s Codex Agent without fragile proxy layers. For anyone running coding agents on a budget, this combination undercuts GPT-4-class pricing by an order of magnitude while delivering competitive Agent benchmark scores.

Summary

The official V4-Flash release posts strong numbers across Terminal Bench 2.1, NL2Repo, DeepSWE, and two DSBench tiers, beating the V4-Pro preview on every Agent metric while holding the preview’s price line. Multimodal input is still absent, but the headline change is native Responses API support on DeepSeek’s own endpoint. Until now, wiring a Chat Completions-only model into OpenAI’s Codex meant either running the older CLI or proxying calls through a format adapter, both of which introduced stability gaps around tool use and streaming. Native Responses API support closes that gap and makes V4-Flash a first-class citizen inside Codex’s Agent loop. A step-by-step guide walks through installing Codex CLI or the ChatGPT desktop app, generating a DeepSeek API key, and using the open-source cc-switch tool to point Codex at the new endpoint with reasoning-effort controls.

Takeaways
V4-Flash official release outperforms the V4-Pro preview on Terminal Bench 2.1, NL2Repo, DeepSWE, DSBench-FullStack, and DSBench-Hard.
Pricing is unchanged from the preview; the model still lacks multimodal input.
DeepSeek’s API now natively speaks the Responses API, eliminating the need for Chat Completions-to-Responses proxies when connecting to Codex.
cc-switch, a 123k-star open-source config switcher, provides a GUI to set DeepSeek as the Codex provider with native Responses format.
Reasoning effort can be set to low, high, or max inside cc-switch’s config.toml section, trading speed and token cost against answer depth.
Codex CLI users verify the switch with /model; desktop users look for a “Custom” model label and a successful test reply.
Conclusions

DeepSeek is deliberately targeting the coding-agent toolchain, not just raw model quality: native Responses API support is a distribution play that makes its cheapest model the path of least resistance inside OpenAI’s own Agent runtime.

The cc-switch tool’s 123k stars suggest a sizable population of developers already routing multiple coding agents through third-party APIs, and DeepSeek’s compatibility makes it an immediate beneficiary of that installed base.

Keeping V4-Flash multimodal-free while pushing Agent benchmarks implies a deliberate product split: Flash handles tool-using code tasks cheaply, while a future Pro or multimodal variant will likely carry the vision workloads at a higher price.

Concepts & terms
Responses API
OpenAI’s newer API format designed for stateful, multi-turn agent interactions with native support for tool calls and streaming. It replaces the older Chat Completions API for agentic workflows.
cc-switch
An open-source configuration manager that lets developers quickly switch between third-party API providers inside tools like Codex and Claude Code, handling model mapping and upstream format selection.
model_reasoning_effort
A parameter exposed by some reasoning models that controls how many inference-time compute cycles the model spends thinking; higher values improve answer quality at the cost of latency and token usage.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗