Native Integration of DeepSeek V4 Flash into Codex

DeepSeek V4 Flash Release

On July 31 DeepSeek officially released the deepseek‑v4‑flash version, adding native support for the Responses API format. This allows Codex to use the model with simple configuration, eliminating the need for local proxy forwarding.

Benchmark Note

Benchmark scores are provided for reference only; actual experience may vary.

Codex Integration

01 Declare Model Metadata

Create a model catalog file at ~/.codex/deepseek-model-catalog.json. This file is referenced when running the /model command to select the deepseek‑v4‑flash model and its inference strength.

https://share-1256249917.cos.ap-chengdu.myqcloud.com/deepseek-model-catalog.json [1]

02 Configure Profile File

It is recommended to use a dedicated TOML file instead of the default ~/.codex/config.toml. The author uses ~/.codex/deepseek.config.toml with the following content:

model = "deepseek-v4-flash"
model_provider = "deepseek"
# preferred_auth_method = "apikey"
# forced_login_method = "api"
model_reasoning_effort = "high"
model_catalog_json = "~/.codex/deepseek-model-catalog.json"

[model_providers.deepseek]
name = "deepseek"
base_url = "https://api.deepseek.com/"
wire_api = "responses"
experimental_bearer_token = "<your DeepSeek API Key>"

The experimental_bearer_token should be filled with the API key obtained from the DeepSeek Platform. The two commented options ( preferred_auth_method and forced_login_method) indicate API‑Key authentication and skipping ChatGPT account login; the author leaves them commented because they also use an official subscription.

03 Start Using the Model

After configuration, navigate to the project directory and run:

cd /path/to/my-project
codex --profile deepseek

The command launches Codex with the DeepSeek V4 Flash model.

Cost and Token Usage

The following table summarizes token consumption and pricing for DeepSeek‑V4‑Flash (0731) versus DeepSeek‑V4‑Pro:

Model Version                | Context Length | Max Output | 1M tokens input (cache hit) | 1M tokens input (cache miss) | 1M tokens output
----------------------------|----------------|------------|----------------------------|-----------------------------|-------------------
DeepSeek‑V4‑Flash‑0731      | 1M             | 384K       | 0.02 CNY                    | 1 CNY                       | 2 CNY
DeepSeek‑V4‑Pro            | 1M             | 384K       | 0.025 CNY                   | 3 CNY                       | 6 CNY

OpenCode Integration

OpenCode offers a simpler path: download the CLI tool and use it directly. It provides a free daily quota and a subscription option (first month $5, second month $10 for the GO plan). For users who only need DeepSeek V4 Flash, usage is effectively unlimited.

Final Thoughts

The author favors DeepSeek among domestic large models because of its low pricing, fast speed, open‑source nature, and lack of hype. With a 1 M context window and about 95 % effective context utilization, DeepSeek V4 Flash is well‑suited for long‑running tasks, unlike Codex CLI combined with GPT models that are limited to roughly 258 K context and require frequent compression.