Backend · Programmer · Artificial Intelligence

DeepSeek-V4-Flash Lands in Codex with One-Click Setup and a Multi-Model Proxy

By cxuanAI · Aug 4, 2026

Read original on juejin.cn ↗ Google Translate ↗ Alt translation

DeepSeek's official integration script is fast but strips Codex of multi-model flexibility. OpenCodex gives that back without modifying the app binary, so developers can benchmark Flash against GPT-5.6-Sol or any other provider inside the same tool they already use.

Summary

DeepSeek-V4-Flash launched with agent performance that beats the V4-Pro Preview, and an official shell script makes integration into Codex a single-command affair on macOS, Linux, and Windows. The script prompts for an API key and sets the model to Custom, but it locks Codex into DeepSeek-only mode, hiding existing projects and sessions.

OpenCodex, a community-maintained local proxy, fixes that limitation. It merges model catalogs from OpenAI, DeepSeek, and other providers into one list Codex can read, then routes requests by provider prefix. Two lines in `~/.codex/config.toml` point Codex at the proxy, and the `ocx` CLI handles install, sync, and teardown.

The result is a native model switcher inside the Codex app and CLI, with the ability to run `codex -m "deepseek/deepseek-v4-flash"` or `codex -m "gpt-5.6-sol"` on the fly. Uninstall or temporary stop commands restore the original configuration cleanly.

Takeaways

— DeepSeek-V4-Flash benchmarks above the V4-Pro Preview and ships with stronger agent capabilities.

— The official integration script requires Codex CLI ≥ 0.144.0 and sets the model to Custom after a single API-key prompt.

— Running the official script hides all existing Codex projects and sessions and removes the ability to switch models.

— OpenCodex is a local proxy that merges model catalogs from multiple providers and injects them into Codex via two config.toml entries.

— With OpenCodex, model switching works natively in the Codex app and CLI, including on-the-fly selection with `codex -m "provider/model"`.

— OpenCodex does not modify the Codex binary; it intercepts requests at `127.0.0.1:10100` and routes them to the correct API endpoint.

— Commands like `ocx stop`, `ocx uninstall`, and `ocx status` provide clean teardown and status checks.

Conclusions

DeepSeek's official one-click script prioritizes a frictionless first run over ongoing workflow compatibility, which creates immediate friction for anyone with existing Codex sessions.

OpenCodex's approach of rewriting only the model catalog URL and base URL in a config file is a lightweight pattern that could apply to any OpenAI-compatible client, not just Codex.

The fact that a community proxy shipped multi-model support within days of Flash's release suggests demand for provider-agnostic tooling is outpacing first-party client features.

Concepts & terms

OpenCodex

A community-maintained local proxy that merges model catalogs from OpenAI, DeepSeek, and other providers into a single list Codex can read, then routes requests to the correct API based on the provider prefix.

Codex model catalog

A JSON file that Codex reads to populate its model picker. OpenCodex replaces the default catalog with a merged one by setting `model_catalog_json` in `~/.codex/config.toml`.

Source: juejin.cn ↗ Google Translate ↗ Backup ↗