Google's Gemini 3.6 Flash uses 17% fewer tokens, a math fix for runaway agent costs

The dirty secret of AI agents isn’t intelligence—it’s economics. Every token a model spits out while reasoning through a task is a micro-transaction, and at thousands of runs per hour, verbosity kills margins. Google’s Gemini 3.6 Flash and 3.5 Flash-Lite, released this week, are a direct shot at fixing this equation, prioritizing output parsimony over parameter flexing.

For teams building background agents, 3.6 Flash is the new workhorse. The headline metric is a 17% drop in output tokens compared to the previous generation, but Google’s developer docs hint at much bigger savings on specific coding benchmarks. On the Datacurve DeepSWE test, token usage reportedly plummeted up to 65%. That’s not an incremental tweak; it’s the difference between a profitable agent pipeline and a cost-overrun disaster. Performance didn’t tank either—MLE Bench scores jumped from 49.7% to 63.9%. Figma’s engineering director, Matt Colyer, confirmed the model is already accelerating design iterations without sacrificing quality.

If 3.6 Flash is for quality reasoning, 3.5 Flash-Lite is the throughput king for grunt work. Clocked at 350 output tokens per second, it’s designed to chew through high-volume document processing and search tasks at a bargain-bin price of $0.3 per million input tokens. Its long-context success rate on Google’s GDM-MRCR v2 test shot from 60.1% to 72.2%, nearly doubling its knowledge work score. The strategy is clear: route simple subagent calls to this cheap, fast model and reserve the heavy compute for complex, multi-step reasoning.

Google also slipped in a specialized variant, Gemini 3.5 Flash Cyber, restricted to vetted partners for vulnerability remediation. It’s a tacit admission that speed matters in security, too. Inside Google’s CodeMender agent, multiple instances of Flash Cyber run in parallel, cross-checking each other’s work before a human signs off. It’s a pragmatic, locked-down approach to a problem—automated scanners outpacing patch teams—that feels more like industrial process control than flashy chatbot demos. The real story here isn’t a model launch; it’s Google selling a bill of materials for building AI that doesn’t hemorrhage cash.