Claude Opus 5.5 and Sonnet 5.5 arrive in Google Antigravity with mid-session switching

Anthropic's newest Claude models land in Google's agent platform

Anthropic's Claude Opus 5.5 and Claude Sonnet 5.5 are now officially available in Google Antigravity, according to a Google AI post on dev.to published October 20. The rollout covers Antigravity 2.0 Desktop, the CLI, and the IDE extensions for VS Code, JetBrains IDEs, and Visual Studio, which puts Anthropic's models alongside Google's own Gemini lineup inside a single agentic coding environment.

The premise of the post is that engineering tasks vary enormously in difficulty, and routing everything through one model is wasteful. Sorting a failed unit test or applying a lint fix does not require the same capability as restructuring a legacy service across thirty files, and Antigravity now gives developers a second model family to divide that work between.

A division of labour between the two models

According to the post, the split is straightforward:

• Claude Opus 5.5 handles the demanding work: deep reasoning, architectural refactoring across many files, and planning that stretches over multiple turns.
• Claude Sonnet 5.5 is the fast, efficient, lower-cost option for routine coding and iterative feature delivery.

Switching models mid-session

The most practical detail is that developers can move between models during a live session, either from a dropdown in the interface or with a command such as /model claude-opus-5-5 in the CLI, and nothing gets reset. Conversation context, open terminal states, and workspace artifacts all survive the switch. That matters because it makes the "right model for the job" idea cheap to act on: a session can start on Sonnet for cost reasons and escalate to Opus only when the task proves harder than expected.

Governance and billing stay on Google Cloud

For enterprises, the post says the Claude models are billed through a single Google Cloud account, and Antigravity's existing security controls, including sandbox policies, MCP server boundaries, and URL allowlists, apply across model families. Adding another vendor's models therefore does not mean standing up a second governance or procurement surface.

Two multi-model workflows

The post also sketches two patterns that become practical once several model families share one agent harness.

Cross-model verification: when the same model both writes code and critiques it, it tends to miss its own mistakes. The suggested approach is to plan and implement a change with Claude Opus 5.5 or Google's Gemini 4 Argon, then hand the diff to a Gemini Pro subagent that checks edge cases against the test suite and design constraints before a human reviewer sees it.

Frontier models paired with flash subagents: mechanical chores such as full-codebase searches or test triage consume latency and budget when run on a frontier model. Antigravity's subagent configuration files, stored under .agents/agents/*.md, let developers declare a lightweight flash model for those jobs. The post gives the example of a test-triager agent that runs test targets, classifies failures by root cause such as compilation errors or timeouts, and returns a structured summary to the parent agent, leaving the expensive models free for reasoning.

Why it matters

Agentic coding tools have so far been largely single-vendor affairs, with each vendor's agent pushing its own models. Anthropic's frontier models shipping inside Google's own product, sold and governed through Google Cloud, is a meaningful break from that pattern. It normalises heterogeneous model routing, where cost, speed, and capability can be chosen per task rather than per subscription, and it makes independent cross-checking between model families a routine option rather than a workaround.

One caveat: everything above comes from a single vendor post, which includes no benchmarks, pricing figures, or third-party evaluation. How well the mid-session switch preserves long context in real projects, and how Opus 5.5 and Sonnet 5.5 actually compare inside Antigravity, will only become clear once developers put the integration to work.