Gemini 3.6 Flash cuts output tokens by 17%

Google shipped Gemini 3.6 Flash on July 21, priced at $1.50 per million input tokens and $7.50 per million output tokens. DeepSWE code precision rose from 37% to 49%, and OSWorld-Verified computer-use accuracy climbed to 83% from 78.4%, the report says. The knowledge cutoff also moved off January 2025, to March 2026, closing a real gap for anyone building on top of it.

The number worth tracking is the 17% drop in output tokens versus 3.5 Flash, per 9to5Google. Benchmarks measure how smart a model is. Token count measures what it costs to run, and that cut applies to every call already running through Flash, no prompt changes required.

Pricing tells a sharper story on the cheap end. Gemini 3.5 Flash-Lite is priced at $0.30 per million input tokens and $2.50 per million output tokens. It beats the full 3 Flash on SWE-Bench Pro, 54.2% versus 49.6%, per 9to5Google. The cheap tier now beats last generation's mid tier on a real coding benchmark. If you route work by cost, the bar for good-enough coding performance dropped in price and rose in capability at the same time.

A 3.5 Flash Cyber variant, restricted to government and trusted partners, handles vulnerability detection and patching, 9to5Google reports. The report doesn't say what qualifies a partner or when that access might widen. The bigger signal sits in that footnote: DeepMind has begun what Google calls its most ambitious pre-training run yet, for Gemini 4. No date, no scale details surfaced. Flash 3 through 3.6 shipped while the actual next generation was already underway behind it.