Google introduced Gemini 3.5 Flash at I/O 2026, delivering output token generation speeds four times faster than rival frontier models. The system achieves a 76.2% score on Terminal-Bench 2.1 while cutting operational costs by more than half compared to competing platforms. Developers and enterprise customers can now access the model across Google's development suite and enterprise products.
Benchmark Performance and Speed Improvements
According to Google DeepMind CTO Koray Kavukcuoglu and Chief Scientist Jeff Dean, Gemini 3.5 Flash is designed to execute complex, multi-step agentic workflows without high latency. In benchmark evaluations, 3.5 Flash outperformed Google's previous Gemini 3.1 Pro model across multiple key coding and reasoning tests.
The system recorded impressive metrics across standardized AI evaluations:
- Terminal-Bench 2.1: 76.2% score for agentic coding execution
- GDPval-AA: 1656 Elo rating for complex problem solving
- MCP Atlas: 83.6% accuracy on tool-use tasks
- CharXiv Reasoning: 84.2% score in multimodal understanding
- Generation Speed: 4 times faster than other frontier models in output tokens per second
Enterprise Deployment and Agentic Workflows
When paired with Google's updated Antigravity harness, 3.5 Flash coordinates collaborative subagents to tackle long-horizon technical tasks. For instance, testing demonstrated that two subagents running in Antigravity synthesized the AlphaZero research paper and coded a fully playable game in six hours.
Several enterprise partners have integrated 3.5 Flash into production operations. Salesforce is embedding the model into Agentforce to manage multi-turn tool calling, while Shopify uses parallel subagents for global merchant growth forecasting. Additionally, Macquarie Bank is piloting the model to accelerate customer onboarding by analyzing documents exceeding 100 pages.
Personal Agents and Safeguard Controls
Google has designated 3.5 Flash as the default model behind the Gemini app and Search AI Mode globally. The technology also powers Gemini Spark, a new 24/7 personal assistant that executes digital tasks under user direction. However, access to Gemini Spark remains limited, rolling out initially to trusted testers before entering Beta for US-based Google AI Ultra subscribers.
Google developed the 3.5 series under its Frontier Safety Framework, introducing upgraded cyber and CBRN mitigations. The release includes interpretability tools designed to evaluate the system's inner reasoning prior to output generation. While 3.5 Flash is generally available today, Google confirmed that its flagship 3.5 Pro model remains in internal testing ahead of a broader release next month.