Gemini 4 Argon reaches AI Pro and Ultra, with a separate gate for standalone Antigravity

Gemini 4 Argon is moving into more paid Google products on October 13. The Verge reports a rollout to Pro and Ultra subscribers in the Gemini app and AI Mode in Search, and to Ultra subscribers in Antigravity. It says availability for Plus and free users remains unclear. This is a reported expansion of access, not evidence that every eligible account already has the model.

The detail to watch is the combination of model, application and subscription. A plan that provides Argon in a chat interface does not, by itself, establish access to Argon in a separate development tool. For a team deciding where to try the model, the useful unit of comparison is the actual workflow: which application will run the task, which account will run it, and whether that account can select Argon there.

Two applications, different access rules

9to5Google distinguishes the routes explicitly. It reports Argon access for Google AI Pro and AI Ultra in the Gemini app and AI Mode, while the Antigravity desktop app and CLI require AI Ultra for the new model. It also reports a coding interface integrated into Gemini for macOS. The standalone-tool requirement should not be treated as proof of identical terms for that embedded interface.

That distinction avoids a misleading shortcut: saying that Antigravity as a whole is restricted to Ultra. The evidence here concerns access to this particular model in the desktop app and command-line tool. It does not establish that every other model or function in those tools has the same restriction. Nor should a screenshot of Argon in the Gemini model picker be used to certify access in a second application.

A practical example is a developer who wants to discuss a change in Gemini and then run an agent in a local project. The first step and the second step have separate access checks. If the account can select Argon for chat but cannot select it in the intended coding environment, the proposed workflow is not yet available as designed. That is an account-and-product observation, not a judgment about the model’s coding ability.

Availability is not a performance result

The Verge describes Google’s claims about reasoning and internal engineering uses, including security work and YouTube infrastructure. Those are vendor claims reported by the outlet. They do not establish the result a different developer will get on a different codebase. This article does not report our own hands-on test, an independent benchmark replication, or a completed rollout across all accounts.

The access change still matters without treating those claims as proven. It gives an eligible user a route to evaluate the model in a named product. A useful evaluation could compare the same bounded task before and after switching models, keeping the input, expected result and review method fixed. Merely obtaining a plausible answer would answer a different question from whether the output satisfies the task’s requirements.

For coding work, an evaluation record should identify the interface and model actually used. A result from chat is not automatically a result from an agent working with a project. The latter comparison may include the files the tool could read, the changes it proposed and the checks applied afterward. These are suggested comparison criteria, not claims that Google mandates a particular testing process or that Argon has passed those checks.

The introductory API rate has an end date

9to5Google reports introductory Gemini API pricing of $2 per million input tokens and $10 per million output tokens through November 30, 2026, followed by $4 and $20 respectively. These are reported API terms, separate from the subscription access rules above. A consumer plan’s inclusion of Argon should not be read as confirmation of a particular API entitlement or API bill.

For the same input and output token volumes, doubling both rates doubles that portion of the charge. That calculation does not predict a complete deployment bill. If the number of tokens changes, the comparison changes too. A developer assessing the introductory offer can record the input and output volumes of a representative task and apply both sets of rates, while labeling any additional costs separately.

That makes the decision concrete without turning launch pricing into a performance claim. One question is whether the desired product route is accessible today. Another is whether the model completes the intended task satisfactorily. A third is what the recorded usage would cost under each rate period. Evidence for one does not settle the others; they can be checked separately before adopting a workflow.

Check for Argon in the exact app you intend to use, then verify that app’s plan requirements before changing your subscription or workflow.