Google Drops Three Gemini Models & Starts Gemini 4 Training

Overview

Google DeepMind released three new Gemini models in a single announcement: Gemini 3.6 Flash , Gemini 3.5 Flash-Lite , and Gemini 3.5 Flash Cyber . The trio is designed to make AI agents running in production faster, smarter, and cheaper. On the same day, Google confirmed it has begun its "most aggressive pre-training ever" for the next flagship, Gemini 4 , which could arrive by year‑end on a typical six‑month cycle.

Gemini 3.6 Flash: Fewer Tokens, Lower Cost, Stronger Benchmarks

Positioned as the new workhorse, 3.6 Flash reduces output token consumption by 17% on average (Artificial Analysis Index) and up to 65% on the DeepSWE coding benchmark compared to 3.5 Flash. It also completes multi‑step tasks with fewer reasoning steps and tool calls, yielding faster execution.

Pricing drops to $1.50 per million input tokens and $7.50 per million output tokens , undercutting 3.5 Flash. Despite being cheaper, 3.6 Flash improves across key benchmarks:

DeepSWE (coding): 37% → 49%

MLE Bench (ML research): 49.7% → 63.9%

OSWorld‑Verified (computer control): 78.4% → 83%

GDPVal‑AA v2 (knowledge work): 1,421 tasks solved, ~70 more than predecessor

Demos show 3.6 Flash orchestrating multiple agents for complex code migration and generating 3D workflows via Gemini Canvas, producing professional‑grade texture‑extraction tools and immersive theme studios.

Gemini 3.5 Flash‑Lite: Speed and Cost for High‑Volume Workloads

Flash‑Lite targets high‑throughput scenarios like massive document processing and agent‑driven search. It delivers 350 tokens/second — the fastest in the 3.5 series — at a rock‑bottom price of $0.30 per million input tokens and $2.50 per million output tokens .

Remarkably, this lightweight model outperforms the larger Gemini 3 Flash on several agent and coding tests:

SWE‑Bench Pro : 54.2% vs 49.6%

OSWorld‑Verified : 74.0% vs 65.1%

The suggested architecture pairs 3.6 Flash as the "brain" (task decomposition) with multiple Flash‑Lite instances as "workers" (parallel execution). An official demo generated 25 production‑ready web designs in seconds. Flash‑Lite is already live in the Gemini App and rolling into Google Search.

Gemini 3.5 Flash Cyber: Specialized Vulnerability Detection

Flash Cyber is a security‑focused model built to find and patch vulnerabilities. Integrated into the CodeMender agent, multiple Cyber agents collaborate to achieve state‑of‑the‑art results on the CyberGym benchmark at lower cost than larger models. The article notes that AI vulnerability discovery now outpaces the ability of existing systems to patch them, making automated remediation critical. Flash Cyber is not yet publicly available.

Gemini 4 Pre‑Training Underway

Google confirmed that pre‑training for Gemini 4 has started, calling it the company's most aggressive training run to date. Based on the historical six‑month cadence, a late‑2024 or early‑2025 release is plausible. The current three models are shipping improvements available today; Gemini 4 remains "futures" for now.

Implications for AI Agent Economics

The release collectively lowers the cost of running AI agents: token usage drops, per‑token prices drop, and per‑agent run cost drops . For developers paying for token consumption and building agent pipelines, these immediate gains are more tangible than the still‑pending flagship.