DeepSeek-V4-Pro Responses Configuration Workflow in Codex

One-sentence takeaway

DeepSeek-V4-Pro can be connected to the Codex CLI, the ChatGPT desktop app, and the VS Code extension through the native Responses API; a single configuration is shared across them, but you should back up and validate config.toml/models.json first.

Use cases

• Suitable tasks: Developers who want to try DeepSeek-V4-Pro at low cost within Codex's existing tool and project workflow.

• Unsuitable tasks: Machines without backups or without a way to manage API keys; remote installation scripts should not be run blindly.

• Applicable model versions: The official documentation lists both deepseek-v4-flash and deepseek-v4-pro; this article uses deepseek-v4-pro.

• Applicable clients, agents, or APIs: Codex CLI, ChatGPT desktop, and the Codex VS Code extension; the communication protocol is the Responses API.

• Recommended reasoning level and parameters: Start with model_reasoning_effort="high"; for complex tasks, compare max, and for simple tasks, compare low.

Ready-to-use content

Core fields for the official manual configuration (put the API key in secure secrets management; do not commit it to the repository):

model = "deepseek-v4-pro"
model_provider = "deepseek"
preferred_auth_method = "apikey"
forced_login_method = "api"
model_reasoning_effort = "high"
model_catalog_json = "~/.codex/models.json"

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

Test/workflow steps

1. Confirm that the Codex CLI or desktop app has been started at least once, and back up the existing ~/.codex/config.toml and model directory.

2. Generate or verify models.json according to the official documentation, confirming that V4-Pro's context, effort, and tool metadata match the current API.

3. In config.toml, set model_provider=deepseek, wire_api="responses", and model_reasoning_effort="high"; if you switch to environment variables or the system keychain for the key, do not copy the plaintext field verbatim.

4. Start with a read-only task and check the Codex startup banner, model selection, Responses requests, and tool calls; only then allow file writes or command execution.

5. After testing, record the model, effort, tool trace, cost, and fallback method; if switching back to ChatGPT login, restart the client as instructed officially and confirm the history grouping.

Raw evidence and data

• DeepSeek says that the Codex CLI, ChatGPT desktop app, and VS Code extension share the same configuration file.

• The official script backs up config.toml, writes models.json, updates the provider configuration, and validates TOML/JSON syntax before writing.

• wire_api="responses" indicates use of the Responses API natively supported by DeepSeek; model_reasoning_effort controls thinking depth and response time.

• The official documentation says that after switching to a third-party API, Codex groups sessions by login method; restoring the original configuration does not delete old records.

Scope and limitations

• This is a configuration workflow, not a code-quality benchmark for DeepSeek-V4-Pro; the Codex harness may change the model's performance.

• The official remote setup script modifies the user's configuration; this article records only verifiable fields and did not run the script. For production integration, review the script, pin its version, and back up first.

• The plaintext experimental_bearer_token example only shows where the field belongs; an actual API key must not be written to Git, logs, or shared screens.

• The model catalog and API support may change; incorrect context/effort metadata can cause client misconfiguration.

Source excerpt or observation (compliance-limited short quote)

DeepSeek describes this integration as “all Codex clients ... share the same configuration file,” but shared configuration does not mean shared session history or identical permissions.