Google Launches Gemini 3.6 Flash, Slashes Token Costs by 65%, and Kicks Off Gemini 4 Training

In a late-night triple-drop, Google DeepMind just released three new Gemini models — including one that cuts token consumption by up to 65% on coding tasks — and simultaneously confirmed it has started its "most aggressive pre-training campaign in history" for Gemini 4. The message is unmistakable: Google is going all-in on making AI agents cheaper, faster, and smarter.

Three Models, One Night

On July 22, 2026, Google unveiled:

• Gemini 3.6 Flash — a next-gen workhorse that dramatically reduces cost and token usage
• Gemini 3.5 Flash-Lite — an ultra-fast, ultra-cheap variant hitting 350 tokens per second
• Gemini 3.5 Flash Cyber — a cybersecurity specialist built exclusively for vulnerability detection and remediation

Gemini 3.6 Flash: The Cost-Killer

The headline feature of 3.6 Flash is brute-force efficiency. According to the Artificial Analysis Index, it uses 17% fewer output tokens than its predecessor 3.5 Flash. On the DeepSWE coding benchmark, savings hit 65% — fewer detours, fewer redundant reasoning steps, fewer unnecessary tool calls per task.

Pricing reflects the efficiency gains: $1.50 per million input tokens and $7.50 per million output tokens — cheaper than 3.5 Flash despite stronger benchmark scores. Key performance jumps include:

Benchmark	3.5 Flash	3.6 Flash
DeepSWE (coding)	37%	49%
MLE Bench (ML research)	49.7%	63.9%
OSWorld-Verified (computer use)	78.4%	83%
GDPVal-AA v2 (knowledge tasks)	—	+70 pts

The model also shines at multi-agent orchestration, demonstrated through complex code migration and real-time 3D workflow generation via Gemini Canvas.

Flash-Lite: Small Model, Big Punch

Gemini 3.5 Flash-Lite hits 350 tokens per second at a staggering $0.30/M input, $2.50/M output — purpose-built for high-volume scenarios like bulk document processing and agentic search.

Most impressively, this lightweight model outperforms its larger sibling Gemini 3 Flash on both SWE-Bench Pro (54.2% vs 49.6%) and OSWorld-Verified (74.0% vs 65.1%). The intended architecture: 3.6 Flash as the "master brain" handling planning, with a fleet of Flash-Lite workers executing at scale.

Flash Cyber: Security-First AI

Gemini 3.5 Flash Cyber tackles a painful reality: AI finds vulnerabilities faster than existing systems can patch them. Integrated into Google's CodeMender platform, it has already set a new state-of-the-art on the CyberGym cybersecurity benchmark — using a lightweight model where competitors need heavyweight ones. It is not publicly available for now.

Gemini 4: The Real Bombshell

Beyond the three launches, Google confirmed it has internally kicked off its most aggressive pre-training run in history, targeting Gemini 4. Industry observers note that, following Google's typical ~6-month training cycle, Gemini 4 could arrive as early as late 2026.

Key takeaway: Google isn't just shipping models — it's reshaping the economics of running AI agents in production. With token costs collapsing, a lightweight model beating its larger predecessor, and Gemini 4 already in the oven, the cost of building on Google's AI stack is dropping faster than almost anyone predicted.