Google Antigravity now includes Claude Opus 5.5 and Claude Sonnet 5.5 for Cloud customers, expanding the model catalog that administrators can enable through the Antigravity model picker. For practitioners, the change means you can match a model’s reasoning depth or latency profile to the specific step in a development or operations workflow, without leaving the Antigravity environment.
What changed in Antigravity
Administrators can turn on the new Claude models and then developers can select them directly from the UI or CLI. The same picker also continues to expose Gemini 3.8 Flash, Gemini 4 Argon, and other Gemini offerings. All models share a unified billing and security boundary, so usage is accounted for consistently across the platform.
Why the new models matter for engineers
Claude Opus 5.5 is positioned for complex, high‑reasoning work, while Claude Sonnet 5.5 targets faster, lower‑cost scenarios. This split mirrors the existing Gemini tiering and gives teams a broader cost‑performance spectrum. In practice, you can run a heavyweight reasoning step with Opus, then hand off repetitive look‑ups or classification to a Flash‑class model, reducing latency and spend.
Architectural patterns enabled by multi‑model engineering
Pattern 1 – Cross‑model verification (LLM‑as‑a‑Judge)
When a loop or graph workflow generates code, using the same model for both creation and review can propagate blind spots. By assigning one model family (e.g., Claude Opus 5.5) to implement and a different family (e.g., Gemini Pro) to validate, you obtain an independent check before committing changes. In Antigravity this can be scripted as an automated sub‑agent or invoked manually via the /model command.
Pattern 2 – Reasoning model paired with Flash subagents
Complex tasks often mix deep reasoning with high‑volume, low‑latency operations such as file searches, log greps, or test classification. Antigravity’s sub‑agent configuration (.agents/agents/*.md) lets you declare a model tier per sub‑agent. A primary session can run on Claude Opus 5.5 for strategic decisions, while sub‑agents marked model: flash offload fast look‑ups to Gemini Flash. Sub‑agents marked model: inherit reuse the active session model for any remaining reasoning.
--- name: test-triager description: Runs unit tests, classifies failures, and returns a concise summary of failing assertions. model: flash tools: - read - edit ---
Similarly, a cross‑model judge can be defined with model: pro to invoke Gemini Pro for a final diff audit.
Operational and security considerations
Enabling third‑party models adds them to the same enterprise billing and security perimeter used for Gemini models. Administrators should verify that the model picker is restricted to authorized users, as the choice of model can affect data residency and compliance footprints. Because sub‑agents execute code on behalf of the primary session, ensure that the tools list aligns with the principle of least privilege – only grant read or edit access that the sub‑agent truly needs.
Related CloudNinjas coverage: Google Cloud.
What This Means For Practitioners
Adopt the new Claude models where you need higher reasoning quality, and keep Sonnet or Gemini Flash for speed‑critical steps. Define sub‑agents with explicit model directives to separate concerns and control costs. Review your Antigravity IAM policies to confirm that only trusted roles can toggle model availability, and monitor usage reports to detect unexpected spend across model families.
