Google Releases Three Gemini Models, Begins 'Most Aggressive' Gemini 4 Training

Google DeepMind Releases Three New Gemini Models

Google DeepMind announced three new Gemini models in a single release, targeting production AI agents with improved speed, cost-efficiency, and specialized capabilities.

Gemini 3.6 Flash: 65% Token Reduction on Coding Benchmarks

The flagship release, Gemini 3.6 Flash, emphasizes token efficiency. According to Artificial Analysis Index, it uses 17% fewer output tokens than its predecessor 3.5 Flash overall, and up to 65% fewer tokens on the DeepSWE coding benchmark. It also requires fewer reasoning steps and tool calls for multi-step tasks, resulting in faster and cheaper execution.

Pricing is reduced to $1.50 per million input tokens and $7.50 per million output tokens, cheaper than 3.5 Flash.

Performance improvements 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 completed, 70+ more than previous generation

Demos show 3.6 Flash orchestrating multiple agents for complex code migration, generating 3D workflows via Gemini Canvas, and acting as an AI interaction designer for immersive theme studios.

Gemini 3.5 Flash-Lite: Faster and Cheaper, Beats Larger Sibling

Gemini 3.5 Flash-Lite prioritizes speed and cost for high-volume scenarios like document processing and agent search. It outputs at 350 tokens per second, the fastest in the 3.5 series, with pricing at $0.30 per million input tokens and $2.50 per million output tokens.

Despite being a lighter model, Flash-Lite outperforms the larger Gemini 3 Flash on several benchmarks:

SWE-Bench Pro: 54.2% vs 49.6%

OSWorld-Verified: 74.0% vs 65.1%

The article suggests a division of labor: 3.6 Flash as the "brain" for task decomposition, and Flash-Lite as "clones" for parallel execution. An official demo shows Flash-Lite generating 25 production-ready web designs in rapid succession. Flash-Lite is already available in the Gemini App and will roll out to Google Search.

Gemini 3.5 Flash Cyber: Specialized Vulnerability Detection

Gemini 3.5 Flash Cyber is a dedicated cybersecurity model for finding and fixing vulnerabilities. Integrated into the CodeMender code security agent, multiple Cyber agents collaborate to achieve state-of-the-art on the CyberGym benchmark at lower cost than larger models.

The article notes an emerging imbalance: AI discovers vulnerabilities faster than existing systems can patch them. Flash Cyber addresses this by providing a lightweight, high-precision model for remediation. However, this model is not yet publicly accessible.

Gemini 4 Pre-training Underway: "Most Aggressive" Training Run

Alongside the three releases, Google confirmed it has started pre-training for Gemini 4, described as the "most aggressive pre-training run in Google's history." Following Google's typical six-month training cycle, Gemini 4 could arrive by year-end.

The article concludes that for developers paying for tokens and building agents, the immediate value lies in reduced costs: lower token consumption, lower prices, and lower per-agent-run expenses. Gemini 3.5 Pro and Gemini 4 remain "futures," while the released models are production-ready today.

Reference: https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-6-flash-3-5-flash-lite-3-5-flash-cyber/?utm_source=tw&utm_medium=social&utm_campaign=og