Google releases Gemini 3.8 Flash and 3.8 Flash Cyber

Background

Gemini is Google's family of large language models, and the Flash variants sit in the lighter tier of that family: models tuned for lower latency and lower cost rather than for the largest possible size. On 13 August 2026, Google released Gemini 3.7 Flash^[2]. The next version followed three weeks later^[1].

The release

On 2 September 2026, Google released Gemini 3.8 Flash and Gemini 3.8 Flash Cyber^[1]. The announcement was written by Tulsee Doshi and Raluca Ada Popa^[1]. The release contained two models rather than one: a general-purpose model and a variant built for security work.

Key details

On HLE-Verified, 3.8 Flash scored 54.9%^[1]. HLE-Verified is an evaluation set; a score is the share of its questions a model answers correctly, so the number can be compared across models, but on its own it says little about everyday tasks.

For 3.8 Flash, Google set introductory pricing at $0.75 per million input tokens and $3.75 per million output tokens, running through 31 December 2026^[1]. Tokens are the units models read and write in — roughly word fragments — and output is priced higher because generating text costs more computation than reading it.

3.8 Flash Cyber is aimed at security work, where it is meant to find and fix software vulnerabilities. Access is restricted: the model was made available to vetted defenders through the Fairwind Program^[1]. The reasoning behind such gating is straightforward — a model that can locate flaws and write patches can also help someone attack a system, so the released version goes only to people whose job is defense.

Two different measurements came with the release. Chrome's security team reported that 3.8 Flash Cyber produced 2.6 times more correct patches for Chrome vulnerabilities than the best commercial models^[1]. Separately, Google reported that 3.8 Flash Cyber scored 47.2% pass@1 on CWE-Bench, compared with 47.8% for a leading frontier model^[1]. pass@1 means the share of tasks solved on the first attempt, with no retries.

What it suggests

The gap between 3.7 Flash and 3.8 Flash was three weeks^[1][2], which says something about how often these releases now arrive.

The two security numbers point in different directions. On CWE-Bench, the public benchmark, 3.8 Flash Cyber came in slightly behind a leading frontier model — 47.2% against 47.8%^[1]. On Chrome's own vulnerability workload, the same model produced 2.6 times as many correct patches as the best commercial models^[1]. A benchmark and a real workload measure different things: a benchmark holds tasks still so that results can be compared, while a production environment has its own mix of code, flaw types and tooling. The two figures are also not like for like — one is a model run against a fixed test set, the other is a count of patches that Chrome's security team judged correct.

What the release shows is a pattern now common among frontier labs: a general model and a hardened, access-controlled version for a high-stakes domain, shipped together, with a price explicitly described as introductory and time-limited^[1].