Google ships Gemini 3.6 Flash and confirms Gemini 4 in pre-training | 17% fewer output tokens, $1.5/$7.5 per million
TL;DR
Google on July 21 released Gemini 3.6 Flash — 17% fewer output tokens vs 3.5 Flash, upgraded coding / knowledge-work / computer-use, knowledge cutoff to March 2026, API $1.5/M input and $7.5/M output. Also shipped 3.5 Flash-Lite and 3.5 Flash Cyber. Confirmed DeepMind has started large-scale pre-training of Gemini 4; Gemini 3.5 Pro in partner testing.
Google on July 21 released Gemini 3.6 Flash — the new model uses 17% fewer output tokens vs Gemini 3.5 Flash and completes multi-step tasks with fewer reasoning steps and tool calls. Improvements across coding, knowledge work and computer use; knowledge cutoff updated to March 2026. API pricing: $1.5 per million input tokens, $7.5 per million output — a 5:1 input/output ratio, the widest gap Google has taken on a Flash-tier release.
Two new variants alongside — Gemini 3.5 Flash-Lite (targeting high-throughput, low-latency workloads, priced for cost-efficiency) and Gemini 3.5 Flash Cyber (for discovering and patching code vulnerabilities — Google's first official vertical-security model variant). 3.6 Flash and 3.5 Flash-Lite are live in the Gemini app and developer platform; Cyber ships in stages.
The real headline is buried at the bottom — Google confirmed Gemini 3.5 Pro is in partner testing and announced that DeepMind has started large-scale pre-training of Gemini 4. This is Google's first public confirmation that Gemini 4 is in pre-training — the market has been guessing for three months, and today's post lands the answer.
Cadence comparison is clear — the past 7 days saw Kimi K3 open-sourced, Qwen3.8-Max, Qwen-Image-3.0, DeepSeek-V4 gray test, four Chinese frontier releases in sequence. Google's play is "small-step-rapid + betting on Gemini 4" — 3.6 Flash + Cyber + Flash-Lite hold API pricing and application territory, while the real weapon Gemini 4 is being held for year-end or 2027 Q1.
Anthropic Opus 4.8 and OpenAI GPT-5.6 Sol lead flagship benchmarks now; Gemini 3.5 Pro entering customer testing means Google is preparing at least one card swap before Q4.
via Google Blog
Two new variants alongside — Gemini 3.5 Flash-Lite (targeting high-throughput, low-latency workloads, priced for cost-efficiency) and Gemini 3.5 Flash Cyber (for discovering and patching code vulnerabilities — Google's first official vertical-security model variant). 3.6 Flash and 3.5 Flash-Lite are live in the Gemini app and developer platform; Cyber ships in stages.
The real headline is buried at the bottom — Google confirmed Gemini 3.5 Pro is in partner testing and announced that DeepMind has started large-scale pre-training of Gemini 4. This is Google's first public confirmation that Gemini 4 is in pre-training — the market has been guessing for three months, and today's post lands the answer.
Cadence comparison is clear — the past 7 days saw Kimi K3 open-sourced, Qwen3.8-Max, Qwen-Image-3.0, DeepSeek-V4 gray test, four Chinese frontier releases in sequence. Google's play is "small-step-rapid + betting on Gemini 4" — 3.6 Flash + Cyber + Flash-Lite hold API pricing and application territory, while the real weapon Gemini 4 is being held for year-end or 2027 Q1.
Anthropic Opus 4.8 and OpenAI GPT-5.6 Sol lead flagship benchmarks now; Gemini 3.5 Pro entering customer testing means Google is preparing at least one card swap before Q4.
via Google Blog
