Google launches Gemini 3.5 Flash with agentic coding focus

Google has released Gemini 3.5 Flash, positioning it as a frontier-level model that matches the intelligence of larger flagship systems while running roughly 4x faster on output tokens. The company claims it outperforms Gemini 3.1 Pro on agentic and coding benchmarks including Terminal-Bench 2.1 (76.2%), GDPval-AA, and MCP Atlas (83.6%), with strong multimodal reasoning scores on CharXiv. A 3.5 Pro variant is in internal use and slated for release next month.

The model is being pushed primarily as an agentic engine, designed to handle long-horizon, multi-step workflows under supervision. Google is pairing it with its Antigravity harness to coordinate subagents for tasks like refactoring legacy codebases to Next.js, generating UI prototypes, and running parallel builder/player loops to develop games. Enterprise partners including Shopify, Macquarie Bank, Salesforce Agentforce, Ramp, Xero, and Databricks are cited as early adopters automating document-heavy or data-intensive workflows.

Gemini 3.5 Flash is now the default model in the Gemini app and Google Search’s AI Mode, and powers the new Gemini Spark personal agent rolling out to trusted testers. Google says the release was developed under its Frontier Safety Framework with strengthened cyber and CBRN mitigations, plus interpretability tooling intended to inspect the model’s reasoning before responses are returned.