Claude Opus 5.5 in Antigravity: Access, Billing and Model Controls

Google announced Claude Opus 5.5 and Sonnet 5.5 availability in Antigravity for Google Cloud customers on October 20. Developers can use the models inside Google's coding environment, but administrators must enable access, and enterprise Claude usage is billed separately from pooled Gemini credits.

For a team already using Antigravity, the useful change is the ability to compare model families without moving every task into a different development tool. The first decision, though, is whether the organization has actually approved that usage and its cost.

Which Antigravity users can use Claude 5.5?

Google's announcement covers Antigravity 2.0, the CLI and IDE extensions. Once an administrator has enabled the models, developers can select Opus 5.5 or Sonnet 5.5 in the model picker.

The broader model documentation also lists both models for paid, non-trial Google AI Pro accounts and Google AI Ultra. That consumer-plan access should not be confused with the enterprise billing path described below.

The same documentation says a model change made during a running turn takes effect after the current execution finishes, unless that execution is cancelled. If a developer switches models to review a change, they should check which model actually produced the review rather than assume the picker retroactively changed an ongoing task.

This is an integration announcement, not the release of a new Claude model. It also does not establish that every Antigravity plan includes identical limits.

Does Claude use the Gemini Enterprise quota?

Claude usage is pay-as-you-go through Cloud Billing, rather than part of the pooled Gemini Enterprise Standard or Plus token quota. Google requires overages to be enabled before third-party models can be used.

The administrator documentation describes an explicit opt-in, acceptance of the provider's terms and a project-level monthly spending limit. It also notes that selecting Claude does not remove Gemini from every background function: lightweight Gemini models can still handle coordination and summaries, and image generation uses a Google model.

There is no single flat “Claude in Antigravity” price in the announcement. Enterprise buyers should check the relevant Cloud Billing SKUs and the configuration attached to their project, rather than copy a direct Anthropic API price into their forecast.

For budgeting, treat the completed engineering task as the unit of work. A migration that produces several planning attempts, reviews and failed test cycles can cost much more than one successful chat response. A cheap individual call is only one part of that calculation.

How administrators enable access

The main controls live under Gemini Enterprise, Settings, then AI developer tools. Administrators manage the permitted models here.

The administrator documentation lists account, licensing and permission prerequisites. Check them before assuming a missing model is a client problem.

These details explain a common rollout problem: a model can appear in a public product announcement while remaining unavailable to a particular employee. Before reinstalling a client, check the account, subscription, organization allowlist and billing configuration.

Keep the rollout small initially. A few representative tasks can reveal whether the chosen model improves accepted output enough to justify its usage. That is a more useful comparison than enabling everything and examining the invoice a month later.

A second model can improve review, but it is not a test suite

One practical use is to ask a different model family to inspect a proposed change. For example, a team could have Claude implement a database migration, then ask Gemini to examine rollback behavior and compatibility with older records.

The review needs a concrete brief. “Does this look good?” invites a broad opinion. “Identify what happens if the migration stops halfway and list the tests that establish recovery” gives the reviewer a narrower job with evidence that an engineer can inspect.

A second model's agreement is not proof that the implementation is correct. Both models may miss the same missing requirement or accept an incorrect assumption supplied in the task. Run the tests, inspect the diff and verify the behavior the user actually requested.

There is also a cost to more review. If the second model merely repeats the first model's explanation, the team has added another billable step without gaining useful evidence. Measure the defects caught and the changes accepted, not the number of agents involved.

What to compare before making it the default

Use a small set of recurring tasks: a bug fix with a reproduction, a multi-file feature and a migration with a rollback requirement. Keep the repository context and acceptance criteria consistent, and compare completion time, human rework and total usage.

That evaluation will not produce a universal winner. It can produce something more useful: a decision about which model earns its place in this team's workflow, and which tasks do not need the more expensive option.

The model picker is becoming easier to change. The engineering discipline around that choice still determines whether a multi-model setup improves the work.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help technical teams turn product updates and engineering decisions into clear stories that developers and founders want to read.

Book a 30-minute call