Gemini CLI Free Tier Shut Down

Fix 1: Switch from Google login to an AI Studio API key

This is the core move and it resolves the "no longer supported" message directly. Create a key at aistudio.google.com/apikey, then:

export GEMINI_API_KEY="your_ai_studio_key"
gemini "say hi"

A working response means you're back in business. To confirm the CLI is actually on the key and not silently falling back to a cached login:

echo "${GEMINI_API_KEY:0:6}..."   # should show key chars, not empty

For a menu, run /auth inside the CLI and pick the API key option instead of "Login with Google." The setting is saved in ~/.gemini/settings.json so you only do it once.

Precedence caveat: the client libraries accept both GEMINI_API_KEY and GOOGLE_API_KEY, and if both are set, GOOGLE_API_KEY wins. If your key looks correct but behaves oddly, check for a stray GOOGLE_API_KEY in your shell profile left over from an old Vertex experiment.

Fix 2: Reset a broken auth state before you blame the key

If you switched keys but the CLI still tries to log in with Google, the old selectedAuthType is cached. Clear it:

# Back up first, then reset the CLI's saved auth choice
mv ~/.gemini/settings.json ~/.gemini/settings.json.bak
gemini   # will prompt /auth again; pick the API key option

This is also the fix for "Please set an Auth method" and for the case where the CLI refuses to switch auth types while an API key is set. A stale settings file causes more "it still doesn't work" reports than any actual key problem.

Fix 3: Enable billing to escape the 429 limit: 0 trap

This is the fix for every "I added a key and still get 429" report. A raw AI Studio key sits on the Free Tier, and a Google One AI Pro subscription does not push paid quota into it. The metric comes back limit: 0, which reads like a bug but is just an unprovisioned key.

Do this once:

• Open AI Studio billing and link a Google Cloud Billing account to the project behind your key.
• Set a real spend cap. A $0 cap keeps you on the Free Tier and reproduces the 429. Set something you're comfortable with — even $10 is enough to move to paid Tier 1.
• Refresh the CLI auth: clear the cache from Fix 2 or re-export the key, then retry.

Tiers escalate with spend: Free, then Tier 1 (billing enabled), then Tier 2 and Tier 3 at higher thresholds, each raising your rate limits. Tier 1 alone carries a per-window spend allowance that's plenty for individual use.

Fix 4: Right-size the model so the paid bill stays small

You're paying per token now, so pick the model deliberately. Standard Gemini API pricing per 1M tokens:

For most terminal work, Flash-Lite is the right default and keeps a heavy day under a dollar. Move up to Flash for real coding and only reach for Pro on the genuinely hard prompts. Re-check exact model IDs at Google's pricing page before you hard-code them into scripts — names move fast.

Fix 5: Use Vertex AI or a Code Assist license

Both paths were untouched by June 18, which makes them the cleanest option if you already have a Google Cloud footprint. For Vertex AI:

export GOOGLE_GENAI_USE_VERTEXAI=true
export GOOGLE_CLOUD_PROJECT="your-project-id"
export GOOGLE_CLOUD_LOCATION="us-central1"
export GOOGLE_APPLICATION_CREDENTIALS="/path/to/service-account.json"
gemini

Vertex uses your Cloud project's billing and IAM, so quota and access are governed by the project — not by a personal Google account that Google can reclassify. If your organization holds a Gemini Code Assist Standard or Enterprise license, the CLI and IDE extensions keep working with no change at all. For teams, this is usually the least fragile answer.

Fix 6: Abstract your model layer so one vendor can't take you down

This is the fix most worth thinking about, because it's the one that protects you from the next shutdown, not just this one.

The pattern: put a thin routing layer between your application and your model providers. One key, one endpoint, and behind it the ability to fail over from Gemini to Claude, GPT, Qwen, or DeepSeek without rewriting your integration. When a single vendor changes its consumer policy, your CLI and your product do not go down with it.

You can run a gateway yourself, use a managed one, or point your tooling at an OpenAI-compatible endpoint that fronts several providers. The specific product matters less than the principle: never let one provider's policy be a single point of failure for a system you rely on.

This is the architectural decision most teams make after a shutdown, not before. The cheaper move is to make it before.