Gemini 3.5: A Frontier AI That Takes Action

Gemini 3.5 was unveiled, with 3.5 Flash as its first model. This model is a frontier-class model focused on agents and coding, serving as the default model worldwide for the Gemini app and AI Mode in Search, and is available in Google Antigravity, Gemini API, AI Studio, Android Studio, Gemini Enterprise Agent Platform, and Gemini Enterprise. Google stated it can handle long-horizon tasks at less than half the cost compared to other frontier models. 3.5 Pro is currently in internal use and is planned for release next month.

Performance was also put front and center. It outperformed Gemini 3.1 Pro on multiple benchmarks, scoring 76.2% on Terminal-Bench 2.1, 1656 Elo on GDPval-AA, 83.6% on MCP Atlas, and 84.2% on CharXiv Reasoning, a multimodal understanding benchmark. Google emphasized that its output-token speed is 4x faster than other frontier models, and it also landed in the top-right quadrant of the Artificial Analysis index.

When combined with the Antigravity harness, multiple sub-agents handle long-horizon workflows in parallel. Examples presented include automatic classification of unstructured assets, summarizing the AlphaZero paper and implementing the game, converting legacy code to Next.js, generating cityscapes, and a self-improvement loop between a builder and a player. Building on its multimodal foundation, it also creates richer web UIs and graphics.

Real-world use cases were also revealed.

• Shopify uses parallel sub-agents to perform long-horizon data analysis and growth forecasting.
• Macquarie Bank reads documents exceeding 100 pages to accelerate customer onboarding.
• Salesforce connects it to Agentforce to automate complex enterprise tasks.
• Ramp handles more sophisticated OCR processing for complex invoices.
• Xero automates vendor verification and information gathering for 1099 tax documents.
• Databricks supports large-scale data diagnostics and problem-solving suggestions.

On the personal side, Gemini Spark was introduced as a personal AI agent built on 3.5 Flash that operates 24 hours a day. In Search as well, it provides a 24/7 information agent and a more dynamic generative UI experience. Rollout to trusted testers began today, and Google AI Ultra subscribers in the U.S. will get beta access next week. At the same time, in line with the Frontier Safety Framework, Google has strengthened cyber and CBRN safeguards and applied new safety training and mitigations, including interpretability tools.