Google Launches Three New Gemini Models, Announces Gemini 4 Pre-training

Google officially released three new models: Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, and Gemini 3.5 Flash Cyber. They are now available on platforms like Google AI Studio, Vertex API, and Gemini App. Based on developer feedback, this update aims to deliver better large-scale agentic applications with higher intelligence, better token efficiency, and lower costs. Additionally, the company revealed that Gemini 4 has entered pre-training, and Elon Musk responded to the launch.

Confirmed

Gemini 3.6 Flash is positioned by the official team as one of the most capable models, focusing on delivering higher quality results with fewer tokens at the same cost. Compared to 3.5 Flash, it is priced lower ($1.50 per million input tokens, $7.00 per million output tokens) and reduces output tokens by up to 17% in complex workflows. It also shows improvements in writing production-level code, chart analysis, and document understanding. Meanwhile, 3.5 Flash-Lite is one of the smallest and fastest models, with an output speed approaching 350 tokens/s, priced the same as the retiring 2.5 Flash. Google also introduced 3.5 Flash Cyber, targeting cybersecurity to compete with Anthropic's Mythos, which is currently limited to government and trusted partners.

Unconfirmed

There is significant divergence regarding benchmark performance and actual cost-effectiveness. Developers like @CounterReady4774 showed strong agentic performance, and @bindureddy called it the "best chat model in the world." However, @bindureddy also noted that its benchmark scores are lower than the previous 3.5 Flash in some tests, and it is more expensive than Grok and Luna. @truecakesnake mentioned its scores are on par with 3.5 Flash based on Artificial Analysis data, and @XFreeze pointed out that its cost per task is higher than Grok 4.5. These controversies suggest the actual ROI might not fully meet official expectations.

Why it matters

Alongside the dense release of new models, Google's Logan Kilpatrick revealed that Gemini 4 has entered its "most ambitious" pre-training phase. According to @kimmonismus, Gemini 4 will be an entirely new foundation model, and the progress so far is exciting. This indicates Google is aggressively pushing to capture the AI market through rapid iteration and specialized models.

2026-07-21 ~ 2026-07-23 · 157 related posts

Full story(5 episodes)→

Primary sources

6 near-duplicate retellings: Acceptable-Debt-294 · BLCNYY · OwariDa · Keano5567 · osanseviero · rseroter