Codex CLI runs on Google Cloud Vertex AI and adds GPT-6.1 Terra

Codex CLI now has Google Cloud Vertex AI built in as a model provider. If your company buys models through Google Cloud, Codex runs there on your gcloud credentials.

Slide text

01 02 03 04 Vertex AI and Terra

openai/codex · release notes rust-v0.165.0 · #56265

Codex CLI runs on Google Cloud Vertex AI

Built-in provider Signs in with gcloud

Vertex AI leaves some Codex features behind. Web search, image generation and custom tools are off there, and so are reasoning summaries. Function tools and remote compaction keep working.

GPT-6.1 Terra is now in the Codex model lists. You can pick it on OpenAI, on Vertex AI, and on Amazon Bedrock, including its global and United States variants.

In the terminal.

The new /presets menu saves prompts or slash commands in ten numbered slots. Press a digit to run one. A preset can also capture your current model, effort and speed.

The model command now sets model, effort and speed on one line, for this session only. Permissions takes a profile the same way. And reasoning effort can change mid-conversation.

The transcript has a keyboard selection mode for copying without the mouse. Move by word, line or page, select whole lines, and search inside the selection.

When a request hits a biological or cybersecurity safety block, the terminal now explains why. It offers safer follow-up prompts, and the one you pick runs in a fork, leaving your original chat alone.

Codex doctor has a network-only mode for when Codex can't reach its model. It checks your configuration, the provider and the Responses WebSocket, and skips the database and MCP scans.

Credentials.

On a remote executor, Bash and Zsh on Linux and macOS now see placeholders instead of your credentials. A proxy swaps the real ones in. Other shells run unmasked, with a warning.

MCP servers behind an enterprise login can now accept an eligible Codex account grant. Or Codex uses a dedicated enterprise sign-in, which now works with OAuth clients that need a secret.

A backslash in a marketplace URL could fool the host check, so Codex rejects it. On macOS, the sandbox guards the git config even when the folder is spelled in capitals. And replaceable sandbox symlinks are refused.

Fixes.

A compaction that fails partway no longer leaves half a summary in your history. A failed background compaction falls back instead of crashing. And resumed threads stop compacting early from counting reasoning twice.

Resume now finds sessions Codex compressed on disk. Resumed threads get back their own service tier, reasoning summary and personality, and the permission settings you accepted.

A missing model now ends the turn with the provider's message instead of retrying. Windows proxies that demand a sign-in get your Windows account. And quit and exit work while the terminal is offline.

If your team already buys models through Google Cloud, I'd switch Codex to Vertex AI first, since it uses credentials you already have. Setup is in the Codex release notes on GitHub.