AI Pulse · Frontier Labs Watch

Google announced it started pre-training Gemini 4 and released three smaller models built for speed and cost, not raw intelligence. The new arrivals are Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber. The last one matters most if you run software or handle sensitive data.

Flash Cyber is tuned for code review and vulnerability detection. It runs cheaper per token than general-purpose models, which means you can scan more code for less money. Google is aiming squarely at OpenAI’s security tools and the niche vendors that charge a premium for threat analysis. If you already pay for code audits or penetration testing, this is a direct alternative that plugs into your workflow without hiring a specialist firm every quarter.

The other two models, Flash and Flash-Lite, trade some accuracy for faster responses and lower bills. That trade works when you need quick answers at scale, like customer support routing or document summarization, where a perfect answer matters less than a fast, good-enough one. Flash-Lite is the smallest and cheapest of the set.

None of this changes what AI can do fundamentally. It changes what it costs to do it. Lower cost per task means you can justify automating things that were borderline before. A security scan that cost five cents per file might now cost two, and suddenly scanning your entire codebase weekly makes sense instead of monthly.

The Gemini 4 pre-training note is a signal, not news you act on today. It tells you Google is still investing heavily and the next capability jump is months away, not years. For now, the practical move is testing whether Flash Cyber catches vulnerabilities your current process misses, and whether the cost difference justifies switching. Tools like the Omni Command Centre let you route tasks to whichever model makes sense for the job, so you’re not locked into one vendor’s pricing when alternatives like this ship.