Google's Gemini 3.5 Flash is built for agents

Google used its I/O 2026 keynote to launch Gemini 3.5, and the first model in the family, 3.5 Flash, is available now. The framing is agents, not chat. Google's pitch is a fast model that can run long, multi-step jobs on its own, calling subagents through its Antigravity harness instead of pausing for a person at every turn.

The numbers Google reports back that framing. 3.5 Flash scores 76.2% on Terminal-Bench 2.1, 83.6% on MCP Atlas, and 84.2% on CharXiv reasoning. Google says it beats its own 3.1 Pro on nearly every benchmark while running about four times faster in output tokens per second, and that a tuned variant is twelve times faster at similar quality. Flash is now the default in the Gemini app and in Search's AI Mode worldwide, with a larger 3.5 Pro due next month to act as a planner that delegates work to Flash as sub-agents.

The speed claim is the part to test first. A model that is both cheaper and faster changes what is practical to run in a loop, and Google is clearly aiming Flash at developers building agents rather than people typing questions into a box. You can read Google's announcement here.

Why it matters

If you build agents, a frontier-level model that is several times faster and cheaper shifts the cost math on long-running loops, but the benchmark and speed figures are Google's own, so test Flash on your own workload before you rebuild a pipeline around it.