Introducing Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber

Source: DeepMind Date: 2026-07-21 URL: https://deepmind.google/blog/introducing-gemini-36-flash-35-flash-lite-and-35-flash-cyber/

Summary

Google shipped its Flash tier GA on 07-21 — three models, all closed-weight, all API-live today. Gemini 3.6 Flash ($1.50/$7.50 per 1M tok) improves coding and knowledge work while cutting output tokens 17% vs 3.5 Flash (up to 65% on DeepSWE); DeepSWE coding 49% (was 37%), MLE-Bench 63.9%, OSWorld-Verified 83.0%. Gemini 3.5 Flash-Lite ($0.30/$2.50, 350 tok/s) targets low-latency agentic throughput — Terminal-Bench 2.1 54% (was 31%), SWE-Bench Pro 54.2%, OSWorld 74.0%. Gemini 3.5 Flash Cyber is a 3.5-Flash fine-tune specialized for “finding and fixing cybersecurity vulnerabilities,” competitive on CyberGym inside the CodeMender multi-agent system, and “exclusively available to governments and trusted partners” via a limited-access pilot. Notably, Gemini 3.5 Pro is still “testing with partners” — Google shipped the Flash tier and the Cyber model while its flagship remains in escrow.

Implications

Closed-clock capability event — the first in ~2 weeks — but the shape is what matters. Google routed around its stalled 3.5 Pro by shipping the cheaper, off-board tier: efficiency (fewer tokens), throughput (Flash-Lite at 350 tok/s), and specialization (Cyber), not a raw-intelligence leap.

• Coding agents / cost curve: Flash-Lite at $0.30/$2.50 with SWE-Bench Pro 54% and Terminal-Bench 54% is a cheap-agentic-throughput play — the economics of running agents, not the ceiling of what they can do.
• Security / dual-use seam: Gemini 3.5 Flash Cyber is Google’s Glasswing/Mythos analogue — a vuln-finding-and-fixing model, gated to governments + trusted partners, deployed in CodeMender. The dual-use cyber model is now a multi-lab, gated-by-default category.
• Voices / the vapor thread: 3.5 Pro stays “testing with partners” — the escrow that won’t clear, now partially routed around rather than resolved.