The Gemini 3.6 Flash release reads as a routine flash-tier refresh. The number that actually changes how you’d deploy it is buried under the model-name soup: output-token usage.

• 💡 3.6 Flash uses 17% fewer output tokens than 3.5 Flash on the Artificial Analysis Index, and up to 65% fewer on DeepSWE. In agentic loops output tokens dominate the bill, so this is a price cut wearing a quality-note costume.
• 📊 Quality moved with it, not against it: DeepSWE 49% vs 37%, MLE-Bench 63.9% vs 49.7%, OSWorld-Verified 83.0% vs 78.4%. Cheaper and better in the same drop is the rare combination.
• ⚡ Flash-Lite is the router tier: ~350 output tokens/sec at $0.30/$2.50 per 1M. That’s what you reach for when a classifier has to make a low-stakes call a million times a day.
• 🔍 Computer use ships as a built-in client-side tool in the API, not a bolt-on — and the OSWorld numbers suggest they mean it for real agentic workflows, not demos.
• ⚠️ Flash Cyber is the odd one out: a vuln-finding model gated to governments and trusted partners through CodeMender, built for multi-agent orchestration. A frontier-competitive model you can’t download.

That gated cyber model is the tell. Domain-specialized LLMs are splitting into a public tier you rent by the token and a restricted tier you have to be cleared to touch. The HN discussion is fixated on the Flash benchmarks; the access model is the more consequential story. How long before every frontier lab ships a security-specialized model only some customers are allowed to run?