Antigravity free model sunset: finish, port, or pay

Google moved the Antigravity removal date for Claude Sonnet 4.6, Claude Opus 4.6, and GPT-OSS-120b from November 2 to November 18, 2026. After that date, Antigravity's Free and Google AI Plus tiers will have only Gemini 3.8 Flash as the selectable agent model. Open the model picker before you plan the next weekend, and see whether Claude is still doing the work. A side project that leans on Claude now needs a scope decision.

You do not need a rescue plan for every experiment. You need a decision rule for the project that has already taken your evenings. The rule is simple: judge it by its one-model ceiling, the point where the project depends on a single free model for its core behavior. When that model leaves, the project has no quiet fallback.

The individual tier costs $0 monthly and provides no limits on Tab completions or Command requests, plus basic weekly rate limits. A weekend project can feel costless until the model you shaped it around is no longer selectable on your tier.

The deadline is a scope decision

Most side projects stay side projects, and that is fine. The problem starts when a free model becomes load-bearing. When the app's main loop depends on Claude Sonnet 4.6 or Claude Opus 4.6, the sunset is a signal. The project has outgrown the lineup it was built on.

Antigravity's prior model lineup included Google Gemini and Claude. OpenAI and Google made an arrangement for OpenAI frontier models to reach Google Cloud, and GPT models were expected to appear in Google Antigravity. Google started making Gemini 4 Argon available in Antigravity, the company's AI coding tool. The free option is narrowing, while paid options keep more models.

Claude Sonnet 5.5 and Claude Opus 5.5 remain in Antigravity for paid Google AI Pro, Ultra, and Enterprise accounts. Finish a toy before the model leaves. Port a project with a real user to what remains free. Pay when the project has already earned revenue or a committed user. Kill it, and keep your evenings, when none of those are true.

The checklist turns a model sunset into a scope decision

Finish the checklist in one sitting. Each item should be testable in a minute, because the point is to force a decision, not to build a migration plan.

List every Claude-dependent workflow in the project, and note which model each one uses.
Mark each workflow as finish before sunset, port to the remaining free model, pay, or kill.
Run the remaining free model on one representative prompt from the hardest workflow.
Set a recheck date before the cutoff, and put it in the same place as the project notes.

One prompt may be doing the real work, so the first item exposes hidden dependencies. Marking each workflow makes the four choices explicit. Testing the remaining free model grounds the port decision in behavior. A recheck date stops the project from drifting into a half-maintained state after the model disappears.

When you mark a workflow as port, test the remaining free model on the task that made the project worth building. If it handles the core loop, the project can stay near zero cost. If it cannot, the project has just told you what it is worth. The free and Google AI Plus tiers receive a weekly quota refresh, while Google AI Pro at $19.99 per month and Ultra receive five-hour quota refreshes up to a weekly limit.

Pay only when the project has earned it

Paying is a promotion when the project can carry it. A weekend project earns a paid model when it has a user, a time-saving workflow, or revenue that covers the subscription. Without one of those, the paid tier is just a way to keep a hobby alive at a price you did not choose.

For an API route instead, Argon's Gemini API price was scheduled to increase from $2 per million input tokens and $10 per million output tokens to $4 per million input tokens and $20 per million output tokens after November 30, 2026. Do not pay to avoid a decision. Pay when the project has already shown it can carry the cost. When it has not shown that, stop spending time on the model question and keep the project small.