Google's Gemini 3.5 Flash brings frontier intelligence to agents

For the past two years, AI progress was measured by how convincingly a chatbot answered your trivia. Google just decided that era is officially over.

At Google I/O, the company launched the Gemini 3.5 family, leading with Gemini 3.5 Flash. It is built explicitly for coding and autonomous agent workflows, and in internal testing, it managed to build an entire operating system from scratch.

Subagents taking over the terminal

Flash has historically meant light and cheap, but 3.5 Flash routinely beats Google's previous top-tier model, Gemini 3.1 Pro. It scored 76.2% on Terminal-Bench 2.1, 83.6% on MCP Atlas, and 84.2% on CharXiv multimodal reasoning.

Speed is the real hook here. The model generates output tokens four times faster than competing frontier models, often at less than half the cost. DeepMind chief technologist Koray Kavukcuoglu also revealed they built an optimized build running 12 times faster with identical output quality.

That rapid output makes multi-agent setups viable. On stage, Google engineer Varun Mohan showed several autonomous subagents simultaneously writing separate code modules inside Antigravity 2.0, Google's new standalone agent IDE, before fusing them into a functioning OS.

Google product director Tulsee Doshi explained the game plan: when Gemini 3.5 Pro rolls out next month, it will act as the high-level planner and orchestrator, spinning up fleets of 3.5 Flash workers to brute-force tool executions. The model can run autonomously for multiple hours, pausing only when it needs human sign-off for critical decisions.

It is already live as the default engine in the Gemini app and Search AI Mode globally. Google is also rolling out Gemini Spark, an always-on personal agent powered by Flash, heading to US Google AI Ultra subscribers next week.

Broader autonomy brings sharper scrutiny. Google is currently battling a lawsuit after a user died by suicide and nearly caused a mass casualty event following weeks of chatting with Gemini last year. With 3.5, Google says it added tighter cyber and CBRN mitigations alongside interpretability tools that inspect model reasoning before serving answers.

Why it matters

Google figured out the unit economics of real-world AI agents.

Having an ultra-expensive flagship model churn through hundreds of bash commands or edit thousands of lines of boilerplate is a commercial dead end. By turning 3.5 Flash into a rapid, dirt-cheap workhorse while saving 3.5 Pro for pure orchestration, Google offers an architecture that businesses can actually afford to scale.

For developers, this is worth testing immediately inside Antigravity or the Gemini API. If your weekly routine involves routine refactoring, data pipeline maintenance, or UI scaffolding, Flash can compress days of tedious labor into minutes. That said, letting it run completely unchaperoned on sensitive codebases remains a risk; its habit of pausing for human permissions is a feature, not a bug.

Across the industry, this shifts the battleground away from chat benchmarks. Competitors will now have to prove their subagents can execute complex, multi-step tool calls without burning through enterprise budgets or hallucinating halfway through a deployment.