A third Flash release in six weeks
Google released Gemini 3.8 Flash on September 2, 2026, alongside a separate, access-controlled variant called Gemini 3.8 Flash Cyber. It's the company's third artificial intelligence (AI) model release in the Flash tier in about six weeks, according to The Register, and it lands as Google positions its cheaper, faster model line as the one that actually gets used day to day, rather than the flagship line most people read benchmark charts about.
Gemini 3.8 Flash is pitched at software engineering, agentic workflows and multi-step reasoning - the kind of long-running task where a model has to keep making decisions over many steps rather than answer a single prompt. Gemini 3.8 Flash Cyber, which replaces the earlier 3.5 Flash Cyber, is built specifically for vulnerability discovery and automated patching, and Google is not handing it out freely.
What Google says it can do
Google's own benchmark claims, published on its blog, are the source for all of the following and should be read as the company's figures rather than independently audited results:
- On DeepSWE v1.1, a long-horizon software engineering benchmark, Google says 3.8 Flash outperforms larger frontier models despite being a smaller, cheaper model.
- It scores 54.9% on HLE-Verified, a benchmark built to be resistant to memorization.
- On CyberGym, Gemini 3.8 Flash Cyber surpasses both its 3.5 Flash Cyber predecessor and larger frontier models Google compared it against.
- In real-world vulnerability discovery testing across 20 programming languages, it clears a 70% success rate.
- On CWE-Bench, a benchmark for patching known weakness classes, it hits 47.2% pass@1 - meaning it produces a correct patch on the first attempt nearly half the time - at what Google describes as significantly lower cost than competing models.
- Google's own Chrome Security team reported that 3.8 Flash Cyber produced 2.6 times more correct patches than the best commercial models they tested it against.
Pricing and where it shows up
Gemini 3.8 Flash costs $0.75 per million input tokens and $3.75 per million output tokens, but that's an introductory rate that expires on December 31, 2026. From January 1, 2027, both prices double, to $1.50 and $7.50. A token is roughly three-quarters of a word, so the increase applies per unit of text processed, not per request.
The model is already live in the Gemini app for Google AI Pro and Ultra subscribers, in Google's AI Mode search feature, and in Google Sheets. Developers can reach it through Google AI Studio, Android Studio and Google Antigravity, and enterprises get it through Gemini Enterprise.
Why the Cyber variant isn't open to everyone
Gemini 3.8 Flash Cyber is only available through a new vetting scheme Google calls the Fairwind Program, which gives prioritized access to government authorities, operators of critical infrastructure, and maintainers of widely used software. Anyone else has to apply through the program rather than simply calling the API.
That restriction is deliberate rather than a rollout delay. A model tuned to find and patch vulnerabilities faster than commercial alternatives is, by the same token, a model that could find those vulnerabilities for an attacker instead of a defender. Google says the release includes safeguards against misuse for chemical, biological, radiological and nuclear (CBRN) purposes as well as cyber-offense, and points to improved resistance to prompt injection attacks as measured by the Gray Swan benchmark, a third-party suite used to test how easily a model can be manipulated by hidden instructions in the content it processes.
What this means in practice
For developers already building on Gemini, 3.8 Flash is a straightforward, cheaper option for coding and agent workloads through the end of the year, with a clear price increase to plan around from January. For security teams, the more interesting move is the Fairwind Program: it's an admission from Google that a model this good at finding bugs needs a gatekeeper, and it sets a precedent other model makers releasing offensive-capable security tools will likely be pushed to follow. Anyone maintaining critical infrastructure or widely used open-source software has a concrete reason to look into applying for access rather than waiting for a public release that may never come.
