Gemini 3.8 Flash, Explained Simply
Google shipped Gemini 3.8 Flash on September 2, 2026 — its newest fast, lower-cost AI model — along with a specialised Cyber variant built for security work. Here's what actually changed and why it matters, without the benchmark jargon.
⚡ Quick facts
- Released: September 2, 2026
- Context window: 1 million tokens in, 64K tokens out
- Beats: 3.7 Flash on every published benchmark, and Claude Opus 5 on three
- Cyber variant: 2.6x more correct security patches than larger rival models
What Gemini 3.8 Flash actually does
It's a multimodal model — meaning it can take in text, images, audio, video, and PDF files — and it's part of a wave of models that also includes fully open-source efforts like India's Sarvam AI. with a huge 1 million token context window (roughly enough to process a long book in one go) and up to 64K tokens of output. Google built it specifically for long-running coding tasks and AI agents that need to work through many steps on their own.
The video trick: watching on demand, not frame-by-frame
Most AI models process video by breaking every single frame into data upfront, which gets expensive fast for long videos. Gemini 3.8 Flash instead uses agentic video understanding — it can navigate a video's timeline and pull just the frames, transcript snippets, or audio it actually needs to answer a question, similar to how a person would scrub through a video looking for a specific moment. The same on-demand approach powers Gemini features that arrived in Android's September update.
How it compares
Google says 3.8 Flash beats its predecessor, 3.7 Flash, on every benchmark it published — and even edges out Anthropic's Claude Opus 5 on three of them. Independent, side-by-side testing will tell the fuller story over time, but on paper this is a meaningful step up for Google's "fast and cheap" model tier.
Gemini 3.8 Flash Cyber: built for security work
Alongside the main model, Google released Gemini 3.8 Flash Cyber, tuned specifically for cybersecurity tasks like finding and patching vulnerabilities. Google's own Chrome Security team reported it produced 2.6 times more correct patches for Chrome vulnerabilities than the best commercial models available — even ones much larger than it.
Pricing and where to use it
Through the API, Gemini 3.8 Flash costs $0.75 per million input tokens and $3.75 per million output tokens — identical to 3.7 Flash's pricing. Google has said both prices are set to double on January 1, 2027, so this is a limited-time rate. It's available now through Google Cloud's Vertex AI and Google AI Studio; it isn't a separate consumer app you'd download, but the underlying model powering other Google AI products.
Speculative Decoding & Video Token Sampling Pipeline
Gemini 3.8 Flash achieves its ultra-low latency and low compute cost through two core architectural mechanisms: a high-efficiency speculative decoding pipeline and sparse temporal video token sampling.
Instead of feeding continuous 30fps or 60fps video frames into the visual encoder—which explodes context consumption—Gemini 3.8 Flash uses an adaptive frame-sampling algorithm that analyzes scene entropy. Static or low-motion segments are compressed into representative keyframe tokens, while high-motion sequences trigger dense sampling. This reduces video token overhead by up to 70% while maintaining temporal reasoning accuracy.
Frequently asked questions
What is Gemini 3.8 Flash?
Google's newest fast, lower-cost multimodal AI model, released September 2, 2026.
What's new?
Better benchmarks than 3.7 Flash, plus on-demand "agentic" video understanding instead of processing every frame.
What is the Cyber variant?
A version tuned for cybersecurity tasks — Google's own security team saw 2.6x more correct patches versus larger rivals.
How much does it cost?
$0.75 per million input tokens, $3.75 per million output tokens via the API — same as 3.7 Flash, for now.