Gemini CLI alternatives: choose by workflow and API control

Gemini CLI is not retired, but its free door is — so "alternatives" splits into two different decisions. The repository is alive: google-gemini/gemini-cli is not archived, carries Apache-2.0 and was last pushed September 21, 2026, and npm stable @google/gemini-cli is 0.60.0, published September 15, 2026, with nightlies still building daily and no deprecation flag (GitHub and npm APIs, checked September 21, 2026). What ended is access. If you are leaving because the login stopped working, you need new access. If you are leaving because you want your own endpoint behind the agent, you need a new client — and that second one is settled by a protocol wall, not by preference.

One disambiguation before any number. A cryptocurrency exchange also called Gemini publishes its own "Gemini API" docs and its own fee schedule, MacPaw sells a macOS duplicate-file finder named Gemini 2, and the gemini:// protocol has its own terminal browsers. None of their prices belong on this page. Google's own names split six ways too — Gemini CLI, Antigravity CLI, Gemini Code Assist, the Gemini app subscriptions, the Gemini API, and the Gemini Enterprise Agent Platform — so every figure below says which one it came from.

Why people are leaving Gemini CLI right now

Google announced on May 19, 2026 that it was transitioning Gemini CLI to Antigravity CLI, stating that "On June 18, 2026, Gemini CLI and Gemini Code Assist IDE extensions will stop serving requests for Google AI Pro and Ultra, as well as those using it free of charge". Google's consumer deprecation page, last updated September 2, 2026, adds that "you can no longer use the Login with Google option to access the IDE extensions or Gemini CLI", that Standard and Enterprise subscriptions "remain unchanged", and that affected users can migrate to the Antigravity family.

Expect one stale source to contradict all of this. The README on main still advertises "60 requests/min and 1,000 requests/day with personal Google account", repeats the free tier twice more, and mentions neither Antigravity nor June 18 anywhere (fetched September 21, 2026). Google's dated deprecation pages are the later word; the README is the trap.

The second trap is a subscription price anchor. Google AI lists a free tier at $0, Plus at $4.99 per month and Pro at $19.99 per month, and Google's May 19, 2026 announcement gives Ultra at $100 and $200 per month — but these are Gemini app subscriptions, and Pro is precisely the tier the transition post names as having stopped being served, alongside Ultra and free Code Assist for individuals. No tier on that subscriptions page is documented as buying Gemini CLI quota. Whether the Ultra figures still stand was not verified on a current pricing page; the subscriptions page does not list Ultra at all.

Staying: what Gemini CLI costs once the free door is shut

Two metered paths remain. The licence path is a per-seat purchase, and Gemini Code Assist quotas (footer dated September 18, 2026) cap it by request count, noting that the daily limits are "aggregated across all interactions with any model version or family (for example, Pro, Flash) used with the Gemini CLI or agent mode" — so one prompt can spend several requests and switching to a cheaper model buys no headroom.

Prices read from Google's Gemini pricing page on September 21, 2026; the page also shows hourly equivalents and notes that subscriptions bill monthly. The key path is per-token instead. Gemini API pricing, last updated September 16, 2026, lists these two among others — and the Flash figures are a dated introductory price that steps up, which is the sort of number that goes stale in a blog post.

That same page lists a Batch API at a 50% cost reduction, and prints batch columns for these models at half the standard rates. Note the context tiering on Pro: an agentic loop re-sends its context on every step, so quoting only the sub-200k figure understates it. Caching, grounding and tool-use line items were not extracted here, so treat these as base token rates rather than a total. Gemini CLI quota covers the request-counting side in more detail.

The wall that decides your shortlist

Gemini CLI does support a custom endpoint. Its in-repo configuration reference documents GOOGLE_GEMINI_BASE_URL as overriding "the default base URL for Gemini API requests (when using gemini-api-key authentication)", requiring HTTPS unless the target is localhost, with a separate GOOGLE_VERTEX_BASE_URL for the path Google has since renamed the Gemini Enterprise Agent Platform. Google's rendered docs site lists no base-URL override at all, so the variable is documented in the repository but not on the published site — real, just easy to miss.

The catch is the wire format. In the released source on main that URL is passed to the @google/genai SDK as httpOptions.baseUrl with vertexai false, so every request leaves as Google's native Gemini protocol, and the auth enum offers no OpenAI-compatible and no Anthropic option. The key travels in an x-goog-api-key header by default; the same file reads GEMINI_API_KEY_AUTH_MECHANISM, and setting it to bearer moves the key to an Authorization header on the gemini-api-key and vertex-ai paths — which changes the header and not the wire format, so it is no escape either. That source also holds an undocumented GATEWAY auth type, selected from the environment whenever GOOGLE_GEMINI_BASE_URL is set — ahead of GEMINI_API_KEY, behind only the explicit Google-login and Vertex switches — and whose key falls back to an empty string. The word "gateway" appears nowhere in the configuration docs, it was read as code rather than as an API contract, and it still speaks the native protocol, so it rescues nothing.

Kunavo serves /v1/chat/completions, /v1/messages and /v1/responses. Re-probed September 21, 2026: api.kunavo.com/v1/models answers 401, while /v1beta/models and /v1beta/models/<model>:generateContent both answer 404. Kunavo is not a valid target for GOOGLE_GEMINI_BASE_URL, and pretending otherwise would cost you an afternoon. Change the client, not the endpoint.

Kunavo serves no embedding, text-to-speech or speech-to-text model either, so a workflow that needs any of those keeps its existing provider for that step.

Antigravity CLI inherits the same wall. Its install and auth documentation says you can run it "with your own Gemini API key instead of a signed-in Google account" and set GOOGLE_GEMINI_BASE_URL to reach "a different Gemini-compatible endpoint" — Gemini-compatible, not OpenAI-compatible. Separately, its plans page states there is currently no support for "Bring-your-own-key or bring-your-own-endpoint for additional rate limits". Those are compatible statements and the qualifier is load-bearing: you may run on your own key, but doing so buys no subscription quota. Its Individual tier is listed at $0 per month (checked September 21, 2026), and no numeric rate limit is printed for any tier. Read that tier's inclusions narrowly: the list under it names model access, unlimited tab completions, unlimited command requests and "Basic weekly rate limits", and does not name the CLI — the only plan that page names the CLI under is the organization plan via Google Cloud. Whether the free tier reaches agy is not something that page answers either way.

Which alternative fits which reason for leaving

Status read from the GitHub REST API and the npm registry on September 21, 2026. One caution on OpenCode: a separate opencode-ai/opencode repository has been archived since September 2025 and continued as Crush, so configuration or prices found there describe Crush's lineage rather than OpenCode's. Compare the two in Crush vs OpenCode, and see Qwen Code pricing for what its routes cost.

What migrates, and what you leave behind

Qwen Code is the shortest hop, so it is the one worth costing precisely. Both sides were read from their own documentation on September 21, 2026: Gemini CLI's configuration reference and extension reference, and Qwen Code's settings and authentication pages. The lineage claim below comes from Qwen Code's own README, which states the project was originally based on Gemini CLI v0.8.2 and stopped syncing with upstream at Qwen Code v0.1.

The Antigravity route has its own migration story and a narrower one than it first reads: its CLI overview says onboarding "supports a one-time import to automatically migrate your existing Gemini CLI extensions, skills, and settings". It names those three things and is silent on chat history, checkpoints and shell history — which is neither a promise nor a denial about them.

For the Qwen Code route, the whole change is one merged settings block. Set the key in your environment, name the base URL at the API's /v1 root rather than the full chat-completions path, and keep GEMINI.md in the context list so your existing project instructions still load:

{
  "context": { "fileName": ["QWEN.md", "GEMINI.md"] },
  "modelProviders": {
    "openai": [
      {
        "id": "claude-sonnet-4-6",
        "name": "Claude Sonnet 4.6 (Kunavo)",
        "baseUrl": "https://api.kunavo.com/v1",
        "envKey": "KUNAVO_API_KEY"
      }
    ]
  },
  "env": { "KUNAVO_API_KEY": "sk-kn-..." },
  "security": { "auth": { "selectedType": "openai" } },
  "model": { "name": "claude-sonnet-4-6" }
}

Rollback is cheap, and worth arranging before you start. None of this touches your Gemini CLI install: ~/.gemini/ is untouched by a Qwen Code setup, and if you still hold a Code Assist Standard or Enterprise licence or a paid Gemini key, gemini keeps working exactly as before. Keep that route funded until one real task has run end to end on the new client. Copy ~/.gemini/settings.json somewhere safe first anyway — it is the file most likely to be edited by hand twice.

A worked estimate for the new client

This is illustrative token arithmetic, not a measured task cost and not a bill ceiling. Assume one agentic session of 10 steps, each sending 60,000 uncached input tokens — the loop re-sends its context every step — and receiving 1,500 output tokens: 600,000 input and 15,000 output in total. No caching, no tool-call overhead and no retries are modelled, and real sessions vary by more than the gaps below. Rates are live Kunavo catalog prices per million tokens.

Leaving the client does not mean leaving the model family: Kunavo's catalog carries Gemini models alongside Claude and GPT ones, so a Qwen Code or OpenCode setup can still call Gemini 3.8 Flash. Check the live catalog for what is listed on the day you read this. Kunavo's catalog amount is a billing floor rather than a cap — when the upstream reports its charge, the bill is the greater of catalog cost and upstream cost times the applicable markup. The minimum top-up is $10 in prepaid credit, a funding minimum rather than a task fee or a subscription; see billing details.

Note which question that table answers. The cheapest listed rate and the lowest cost to finish your task are different claims: a model that needs three attempts at a low rate can cost more than one that lands it once. Shortlist on rate, then measure one bounded task on your own repository and read what your account actually recorded for it.

Set it up and check the first task

Kunavo publishes a setup reference for the closest hop. That is a published configuration page, not a compatibility test — Qwen Code has not been runtime-tested here against Kunavo's endpoint, and neither has any other client on this page. Start at the Qwen Code integration guide, or the OpenCode, Crush, Cline and Aider pages if one of those fits your reason for leaving better. Create a Kunavo account when you are ready to fund a key.

Still deciding whether to switch clients at all? Gemini CLI vs Claude Code compares the workflows directly, and the OpenCode provider comparison covers the same routing question from the other end.

FAQ

Repository and npm status, Google's deprecation and pricing pages, Antigravity's install and plans documentation, and Qwen Code's own configuration docs all checked September 21, 2026; the Kunavo endpoint probe was re-run the same day. Nothing on this page was runtime-tested: no CLI was installed and no session was run against any endpoint. Kunavo token rates come from the live catalog, and the dollar example is illustrative token arithmetic.