DeepSeek V4 Now Connects Directly to Codex, Moving Risk From Protocol Translation to Configuration
Crafting seamless user experiences with a passion for headless CMS, Vercel deployments, and Cloudflare optimization. I'm a Full Stack Developer with expertise in building modern web applications that are blazing fast, secure, and scalable. Let's connect and discuss how I can help you elevate your next project!
DeepSeek V4 Codex integration now covers both V4 Flash and V4 Pro through a native Responses API. DeepSeek added Flash support on July 31, 2026, then extended the same path to the production V4 Pro release on August 13. The local translation proxy can disappear. Shared configuration, plaintext credentials, model economics, and recovery testing now deserve the scrutiny.
That update changes the selection problem. A week ago, Flash was the only production V4 model with documented Codex support. Both models now use the same official installer and model catalog. Teams are choosing between cost and task difficulty rather than protocol compatibility.
Why native Responses support changes operations
Codex custom providers now speak only the Responses wire protocol. OpenAI's source rejects wire_api = "chat" and directs users to responses. Before DeepSeek exposed a native endpoint, a local service had to translate Codex requests into Chat Completions and convert the stream back again.
Before: Codex -> local protocol proxy -> DeepSeek Chat Completions
Now: Codex -> DeepSeek Responses API
Removing the proxy gives failures a cleaner owner. A broken stream, malformed tool call, or lost reasoning item no longer has to be traced through a second event model. The API key also stops passing through an extra service. None of that improves a benchmark score, but it affects every long-running coding task.
Flash and Pro now share the integration boundary
DeepSeek's current Codex catalog gives both models a 1,048,576-token maximum context and up to 384,000 output tokens. Both accept text input and support low, high, or max reasoning effort. The catalog sets Codex 0.144.0 as the minimum client version.
| Item | DeepSeek V4 Flash 0731 | DeepSeek V4 Pro 0813 |
|---|---|---|
| Terminal Bench 2.1 | 82.7 | 87.9 |
| DeepSWE | 54.4 | 62.7 |
| Toolathlon-Verified | 70.3 | 74.1 |
| Current cached input per 1M tokens | $0.0028 | $0.003625 |
| Current uncached input per 1M tokens | $0.14 | $0.435 |
| Current output per 1M tokens | $0.28 | $0.87 |
| API concurrency limit | 2,500 | 500 |
These are vendor-published evaluations. The Flash model card says its public code-agent tests used the unreleased DeepSeek Harness in minimal mode, max reasoning effort, temperature = 1.0, and top_p = 0.95. Treat the scores as admission criteria for a pilot. They cannot replace regression tasks from your own repositories.
The pricing model changes at 16:00 UTC on August 16, or midnight in Beijing and Taipei on August 17. DeepSeek will charge peak and off-peak rates, with off-peak set at half of peak. Peak windows are 01:00-04:00 UTC and 06:00-10:00 UTC.
| Model and period | Cached input per 1M | Uncached input per 1M | Output per 1M |
|---|---|---|---|
| Flash off-peak | $0.007 | $0.22 | $0.66 |
| Flash peak | $0.014 | $0.44 | $1.32 |
| Pro off-peak | $0.022 | $0.66 | $1.98 |
| Pro peak | $0.044 | $1.32 | $3.96 |
Scheduling becomes part of cost control for batchable work. Flash is the sensible first candidate for high-volume reading, search, and low-risk edits. Pro fits architecture decisions, cross-module debugging, and final checks where a bad answer costs more. Your own success rate and retry bill should decide the final split.
Keep installation recoverable
Codex CLI, the ChatGPT desktop app, and the Codex extension for VS Code share ~/.codex. A provider change reaches all three clients. Quit them fully, make sure Codex has run at least once, and preserve the current configuration before installing.
For an isolated pilot, point the official script at a separate directory.
export DEEPSEEK_CODEX_TEST_HOME="$HOME/.codex-deepseek-test"
mkdir -p "$DEEPSEEK_CODEX_TEST_HOME"
CODEX_HOME="$DEEPSEEK_CODEX_TEST_HOME" \
bash <(curl -fsSL https://cdn.deepseek.com/api-docs/codex-deepseek-setup-en.sh)
To change the existing macOS or Linux environment, use DeepSeek's documented command:
bash <(curl -fsSL https://cdn.deepseek.com/api-docs/codex-deepseek-setup-en.sh)
Windows PowerShell uses:
irm https://cdn.deepseek.com/api-docs/codex-deepseek-setup-en.ps1 | iex
The current menu assigns 1 to Flash, 2 to Pro, and 3 to restore. On first install, the script backs up config.toml, writes a models.json catalog containing both models, adds the DeepSeek provider, and validates TOML and JSON before replacing files. A later run can switch only the model or restore the pre-install state.
The script can read an API key from DEEPSEEK_API_KEY, which avoids an interactive prompt. Keep that export out of version-controlled project files. The official setup still writes the credential into experimental_bearer_token, leaving it as plaintext in config.toml. Teams that prohibit plaintext secrets on disk should stop at an isolated review rather than applying the installer to a primary workstation.
Manual setup needs the full model catalog
Teams that review every field can copy the complete models.json block from DeepSeek's integration page. They can then add the matching provider configuration. This excerpt shows the core fields and does not replace the catalog.
model = "deepseek-v4-flash"
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>"
Switching to Pro changes the top-level model field:
model = "deepseek-v4-pro"
The official catalog also defines tool formats, reasoning levels, compaction thresholds, parallel tool calls, and base instructions. A reduced configuration can return ordinary text while failing under long context or repeated tool use.
Acceptance testing must prove transport, tools, and recovery
The desktop label Custom is not a failure signal. DeepSeek says macOS displays Custom, while Windows may show either Custom or the model name. A useful acceptance test has four layers:
- Configuration: verify the selected model,
wire_api = "responses", and a parseable model catalog. - Transport: find the matching request and timestamp in DeepSeek Platform usage records.
- Agent behavior: run a small task that must search, read a file, edit code, and execute a test.
- Recovery: choose the installer's restore option, restart each client, and confirm that the ChatGPT login and original session group return.
Do not ask the model who it is as a provider test. DeepSeek's catalog gives the model Codex base instructions, so an identity answer can merely repeat the prompt. Billing records and request timestamps are stronger evidence.
Treat provider switching as a configuration migration
DeepSeek's installer preserves MCP servers and project trust settings while removing fields that conflict with its model catalog. Restore copies the pre-install config.toml back into place. Manual edits made after installation can disappear with that restore.
Third-party switchers can manage several snapshots, but they gain permission to read and write credentials and configuration. Review where they store keys, which files they change, whether they show a diff, and whether removal produces a complete recovery.
My default is deliberately conservative: prove configuration, transport, tools, cost, and recovery in an isolated CODEX_HOME before touching the primary environment. Native Responses removes the hardest protocol layer to observe. The remaining risks are visible enough to test.
Frequently asked questions
Can DeepSeek V4 Pro connect to Codex now?
Yes. DeepSeek released the production V4 Pro model on August 13, 2026, with native Responses API and Codex support. The official installer now switches between Flash and Pro.
Should a team choose Flash or Pro?
Start Flash on high-volume, lower-risk work. Reserve Pro for architecture, difficult debugging, and final verification. Rebalance with your own completion rate, retries, and total token cost.
Does Custom in the desktop app mean setup failed?
No. Confirm the startup model, DeepSeek usage records, and a complete agent task. The interface label is only one signal.
Are old conversations deleted after switching?
DeepSeek says Codex groups sessions by login method. Restoring the original configuration and restarting the clients should reveal the subscription session group again.
Can I set only the base URL and model name?
That is unsafe for a production pilot. Codex needs the Responses protocol and model metadata. Use the official installer or create the complete official catalog with matching provider settings.
Sources
- DeepSeek: Integrate with Codex
- DeepSeek: Models and Pricing
- DeepSeek: August 13, 2026 V4 Pro update
- Official DeepSeek V4 Flash 0731 model card
- OpenAI Codex model-provider source
Author Insight
The hardest failures in agent deployments are often the ones with no clean owner. Direct protocol support removes one source of invisible state, then puts credential handling, shared configuration, and recovery back where they belong: inside the engineering acceptance test.
