Gemini CLI v0.62: gemini-3.8-flash and gemini-3.5-flash-lite added

Originally published at https://aicoding-guide.com.

Gemini CLI v0.62.0 is mostly fixes with one model addition in the middle of them.

The thing to know: gemini-3.8-flash and gemini-3.5-flash-lite are now available. Alongside that, a fix retains the OAuth refresh token on refresh, and several fixes land on Windows shell execution.

One caveat about the source: the v0.62.0 release notes are a list of PR titles with no per-change descriptions. This article sticks to what the PR and the official configuration reference confirm, and says so where confirmation wasn't possible.

Key point
What you will learn

• The two new model IDs and which tier each leads
• Setting a model with model.name in settings.json
• The OAuth, Windows and proxy fixes in this release

The two new models

PR #29443 adds both as the latest GA release in their respective tiers.

Model ID	Tier	Position
gemini-3.8-flash	flash	Latest GA in the Flash tier
gemini-3.5-flash-lite	flash-lite	Latest GA in the Flash Lite tier

The same PR promotes gemini-3.5-flash and gemini-3.1-flash-lite to base tier, with the two new models established as the "latest" alternatives. Existing model selections don't disappear — they move within the hierarchy.

Both IDs are also confirmable in the official configuration reference as of October 1, 2026: gemini-3.8-flash appears among the modelConfigs aliases, and gemini-3.5-flash-lite among the model definitions.

Whether you can pick them depends on your auth type
Per the PR description, access is gated behind the LATEST_FLASH_GA_LAUNCHED and LATEST_FLASH_LITE_GA_LAUNCHED experiment flags, with immediate access for non-experiment auth types such as USE_GEMINI, USE_VERTEX_AI and GATEWAY. That gating mechanism could not be confirmed in the official documentation. If the models don't show up for you, a staged rollout is the likely reason.

Setting the model in settings.json

Per the configuration reference, the model used for conversations is set with model.name.

{
  "model": {
    "name": "gemini-3.8-flash"
  }
}

The model object documents these keys:

Key	Default	What the reference says
model.name	unset	The Gemini model to use for conversations
model.maxSessionTurns	-1	Maximum user/model/tool turns to keep in a session; -1 means unlimited
model.compressionThreshold	0.5	The fraction of context usage at which to trigger context compression

Settings files load in this order, each overriding the one before it:

1. System defaults: /etc/gemini-cli/system-defaults.json on Linux, C:\ProgramData\gemini-cli\system-defaults.json on Windows, /Library/Application Support/GeminiCli/system-defaults.json on macOS
2. User settings: ~/.gemini/settings.json
3. Project settings: .gemini/settings.json
4. System overrides: /etc/gemini-cli/settings.json, or the OS equivalent

For finer control, modelConfigs holds aliases — "named presets for model configs" that can inherit from one another with extends — and modelDefinitions, which carries each model's metadata: tier (flash-lite, flash, pro, custom), family (gemini-3, gemini-2.5, gemma-4, auto), isPreview, and features.

For the project context file, see Setting up GEMINI.md in Gemini CLI.

Authentication and login persistence

Two authentication fixes are in the list:

• "retain oauth refresh token on refresh and make credential deletion idempotent" (#29339)
• "update auth error documentation link to valid anchor and add fallback" (#29377)

The first reads as closing a path where the refresh token was lost during a token refresh. It coincides with reports since v0.60.0 of sandboxed sessions not keeping login state, but the release notes carry no description, so whether those reports are resolved could not be confirmed. If you are re-authenticating on every launch, v0.62.0 is worth trying first.

On the MCP side, "format MCP tool call titles as structured signatures and segregate explanations" (#29341) changes how tool calls are presented.

Windows, proxy and the rest

The remaining fixes concern the execution environment.

Area	What the PR title says
PTY	Improve file descriptor cleanup and execution lifecycle management (#29340)
ConPTY (Windows)	Synchronize process exit lifecycle and harden PTY output finalization (#29379)
Terminal	Improve terminal buffer memory management and format Windows diagnostic paths (#29380)
Proxy	Normalize proxy-agent esbuild interop for environment proxy resolution (#29401)
VS Code companion	Preserve terminal focus when closing diff tabs (#29378)
ACP mode	Emit tool_call update prior to request_permission (#29439)
UI	Guard against negative layout dimensions in border rendering (#29347)
Logs	Suppress uncaught AbortError logs during request cancellation (#29343)

If shell command execution has been unreliable on Windows, the three PTY and ConPTY fixes are the ones to look at; if you couldn't connect through a proxy, #29401 is yours. For all of them, the release notes give no description, so exactly which symptoms clear could not be confirmed.

Summary

• The headline in v0.62.0 is gemini-3.8-flash and gemini-3.5-flash-lite, both confirmable in the official configuration reference
• gemini-3.5-flash and gemini-3.1-flash-lite move to base tier, with the new pair as the latest
• Set the model with model.name in settings.json; user settings live at ~/.gemini/settings.json
• A fix retains the OAuth refresh token on refresh — worth taking if your login keeps dropping
• If the new models don't appear, a staged rollout is the likely cause, though that mechanism isn't in the official docs