Overview
Gemini Composite V2 is a voice-optimized composite model for enterprise conversational agents in Google Cloud CX Agent Studio. It became generally available in CX Agent Studio on October 22, 2026; that launch status does not establish its availability in other Google Cloud products. (docs.cloud.google.com)
Capabilities and uses
The model combines listener, thinking and speaker components. It uses the same listener and thinking models as Composite V1, but upgrades the speaker for more natural pacing, tone and expression. Inline tags can control pauses, breaths, laughter and tone shifts. Google also reports improved spoken accuracy for alphanumeric values such as tracking numbers and phone numbers. It is intended for enterprise voice agents handling customer conversations and backend tool workflows. (docs.cloud.google.com)
Limitations and deployment considerations
- Long speaking-style prompts can cause voice drift or conflict with inline tags. Google recommends starting with an empty style or a short speaker description.
- For reliable spoken numbers, dates, currency and identifiers, Google recommends explicit text-normalization rules. Test agents before deployment, particularly when migrating from Composite V1.
- The cited announcement and model documentation do not specify a model context window or separate model-specific input and output token prices. (docs.cloud.google.com)
Pricing
As of November 15, 2026, separate input and output token prices for gemini-composite-v2 were not verified. The official CX Agent Studio service pricing page instead lists USD $0.50 per voice session, with usage beyond the first 300 seconds charged at USD $0.0025 per second. This is service-level session pricing, not a model-specific per-token quote; it does not specify separate input and output charges. If a session starts as chat and switches to voice, the entire session is billed as one voice session. (cloud.google.com)