Gemini CLI 0.67.0 adds models command, sandbox IDE diffs

Google promoted Gemini CLI 0.67.0 to the stable channel on November 4, 2026. The features first appeared in the 0.67.0 preview on October 27. If you skipped the preview, three changes are worth knowing about: a command that lists models, IDE diffs that now work when the agent runs in a Docker, Podman or gVisor sandbox, and fewer confirmation prompts for shell commands.

Release facts

List models from scripts

PR 29880 adds gemini models (also gemini models list) and a top-level --list-models flag. With -o json you get an array with modelId, displayName, description and contextLimits (input and output tokens). The flag exits before sandbox or TUI startup and skips the auth refresh, so credential log lines no longer end up in stdout when you pipe the output to jq.

IDE diffs inside a sandbox

Before this release, /ide enable, open-file context and native diffs did not work when the CLI ran in a container sandbox (GEMINI_SANDBOX=docker, podman or runsc) and the editor ran on the host. PR 29858 fixes this without passing the IDE token into the container. The host launcher relays the session over docker exec -i or podman exec -i. It only allows the openDiff and closeDiff tools, and only for files inside the IDE workspace folders. The sandbox image only needs node.

Fewer shell prompts, same checks on fetched content

PR 29833 addresses repeated confirmation prompts, which users reported even in YOLO mode. Flags that appear only in the output of the agent's own shell commands, such as git diff or a test runner's usage text, no longer count as untrusted input. Workspace paths are no longer reported, and an approval now covers the same flags for the rest of the session. Flags that come from web_fetch, web_search or MCP output still ask for confirmation.

Fixes to check if you script it

• Spending caps: when an AI Studio project hits its monthly spending cap, the CLI used to retry and show "We are currently experiencing high demand." It now stops and shows the API's spending-cap message (PR 29961).
• Cancellation in CI: in non-TTY runs (Docker, CI, scripts), SIGINT and SIGTERM now cancel the run, queued tool calls are not executed, and the process exits with code 130 (PR 29849).
• File writes: the atomic write added earlier silently overwrote read-only files and split hard links. Writes to a 0444 file fail with EACCES again, and files with more than one hard link are written in place (PR 29749).
• Settings migration: ${VAR} placeholders survive the deprecated-settings migration, and Claude hook timeouts are converted from seconds to milliseconds when imported.

Questions

Is 0.67.0 a new Gemini model? No. It is a release of the CLI client. Which models you can use still depends on how you sign in.

How do I see which models my account can use? Run gemini models, or gemini --list-models -o json for machine-readable output.

Do I need to change my sandbox image for IDE diffs? No, according to the PR. The image only needs node, and the bridge does not depend on container networking.

The previous stable release is covered in Gemini CLI 0.66. If quotas are the reason you are looking around, see Gemini CLI alternatives. More tools are in the coding agents category.