Google Cloud adds Claude models to Antigravity tools

Google Cloud has added Claude Opus 5.5 and Claude Sonnet 5.5 to Google Antigravity, making the models available to Google Cloud customers using its AI coding tools.

Developers can choose Anthropic's Claude models from the model picker in Antigravity once an administrator enables them. They are supported across Antigravity 2.0, the Antigravity command-line interface, and IDE extensions for VS Code, JetBrains IntelliJ, and Visual Studio.

Administrators can also select Google's own Gemini models, including Gemini 3.8 Flash and Gemini 4 Argon, in the same environment. Google says this lets teams use different model families for different software tasks while keeping billing and security controls within the same organisational setup.

Model choice

Claude Opus 5.5 is aimed at more complex work, while Claude Sonnet 5.5 is positioned as a faster, lower-cost option. Developers can switch the active model between turns in the same conversation thread without losing workspace state, open artefacts, or terminal context.

The addition broadens the mix of third-party and first-party models available in Antigravity, Google's environment for software engineering tasks. It also reflects a wider industry trend of cloud providers giving customers access to several AI models rather than requiring them to use a single in-house system.

Google outlined two workflow patterns enabled by having multiple model families in the same agent framework. One is cross-model verification, in which one model writes or plans code and another reviews the output before the work moves to the next stage.

Under this approach, a developer might use Claude Opus 5.5 or Gemini 4 Argon to plan and implement a change across several files, then switch to another model family to critique the implementation against tests and requirements. This can be done interactively or through automated subagents.

Workflow patterns

The second pattern combines reasoning-heavy models with lighter-weight subagents for repetitive tasks. Complex software jobs often involve a small number of decisions that require stronger reasoning, alongside many routine actions such as searching files, checking logs, reviewing lint output, or classifying test failures.

Antigravity includes a subagent customisation system that lets users assign different model tiers to specific subagents. Those tiers include inherit, flash, and pro, allowing a session running on a more advanced model to delegate narrower tasks to lighter models.

Google gave the example of a log-triage subagent assigned to a flash model to run unit tests, classify failures, and return a concise summary of failed assertions. It also described a separate cross-model review subagent assigned to a pro model to audit code differences against user requirements and check for regressions, missing edge-case tests, and security issues.

The setup is designed to divide labour between models, with lower-cost models handling high-volume checks and more advanced models reserved for coding or review steps that need deeper reasoning. Google argues that such combinations can help teams manage both response time and spending.

Billing rules

Google also explained how billing works for partner models in the service. For customers using Gemini for business, usage of third-party partner models does not come from the monthly pooled token quota attached to Gemini Enterprise Standard and Plus editions.

Instead, usage of those partner models is billed on a pay-as-you-go basis through an existing Google Cloud billing account, which means overages must be enabled. Administrators can configure overage billing, spending limits, and authorisation for partner models through the Google Cloud Console.

Customers using Antigravity through the Gemini platform API can also enable Claude models in Model Garden after completing Anthropic's standard terms. Developers need the latest version of Antigravity 2.0, the CLI, or the IDE extensions to choose the newly added models.

Google added that lightweight first-party Gemini models will continue to run in the background for high-frequency client-side coordination tasks such as context management and summaries, and that they consume first-party tokens and can incur charges.