Claude Opus 5.5 lands in Google Antigravity, billed pay-as-you-go

Google says Claude Opus 5.5 and Claude Sonnet 5.5 are now available in Antigravity, its agentic coding environment, for Google Cloud customers. The key detail is the bill: according to the announcement, Claude usage is pay-as-you-go, outside the pooled Gemini token quota.

Claude joins the model picker, once an admin says yes

Google describes Opus 5.5 as the model for complex work and Sonnet 5.5 as a faster, lower-cost complement. Developers do not get them automatically. An administrator must first enable them, after which they appear in the model picker. Admins can also choose Google’s own models, including Gemini 3.8 Flash and Gemini 4 Argon.

The models work across Antigravity 2.0, the Antigravity CLI and the recently released IDE extensions for VS Code, JetBrains IntelliJ and Visual Studio. Google pitches the setup as one enterprise billing and security boundary. A conversation can switch models between turns while keeping workspace state, open artifacts and terminal context, according to the post. Developers need the latest version of the tool and can also run /model in the CLI.

Setup differs by route. Administrators on Gemini for business configure overage billing, spend limits and partner-model authorization under “Gemini Enterprise” > “Settings” > “AI developer tools” in the Google Cloud Console. Those using Antigravity through the Gemini platform API enable Claude in Model Garden and accept Anthropic’s standard terms. Google’s post does not list regions, per-token prices or usage caps for the Claude models.

Two patterns for mixing model families

The post spends much of its length on workflows. The first is cross-model verification: one model family plans and implements a change, and a different family reviews it against the tests and requirements before the work moves on. Google’s reasoning is that a model reviewing its own output can repeat its own blind spots.

The second pairs a strong reasoning model with cheaper subagents. Subagents are defined in .agents/agents/*.md files, where a model field accepts inherit, flash or pro. Google’s example sends test triage to a Flash-tier test-triager and final diff review to a cross-model-judge on a Pro-tier Gemini model. Note that the Claude models themselves are not among the tiers named in that field; the examples use Claude for the main session and Gemini for the side jobs.

Prices that make the separate bill easier to judge

The announcement gives no figures, so the cost has to be estimated from outside. According to the pricing tracker UsagePricing, Opus 5.5 launched on October 5, 2026 at $4 per million input tokens and $20 per million output tokens, down from $5 and $25 for Opus 5, which stays on the rate card. That is a 20% cut. Third-party trackers such as Apidog list Sonnet 5.5 at $2 input and $10 output per million tokens. These are secondary sources, not Anthropic’s price page, and Google’s own Cloud rates may differ.

The comparison matters because of how Google structures the bill. Gemini Enterprise Standard and Plus customers get a pooled monthly token quota, but Claude does not draw on it. Every Claude token therefore lands as an extra charge on the Cloud Billing account, which is why Google tells admins to set spend limits first. Google also notes that lightweight first-party Gemini models keep running in the background for context management and summaries, and those tokens can incur charges as well. A team using Claude for the main session is, in effect, paying twice: for Claude’s tokens and for Gemini’s housekeeping.

The Flash-subagent pattern is the obvious way to contain that. Sending high-volume searching and classification to the cheapest tier keeps the expensive model for the few decisions that need it, though Google gives no savings figure.

What the post leaves open

The claims about better code quality and lower latency are Google’s and come without benchmarks or customer data. The post does not say whether Claude features such as the structured outputs recently added to Claude on Amazon Bedrock are available here, how Claude’s tokens are metered inside agent loops, or which regions are covered. For Google, the addition fits its stated aim of an open model ecosystem; for customers, the practical test will be the invoice.