Introducing Gemini 3.8 Flash and 3.8 Flash Cyber

What it claims

Gemini 3.8 Flash is Google DeepMind's third Flash release in six weeks, three weeks after 3.7 Flash. It keeps 3.7's price ($0.75 input / $3.75 output per million tokens), but the page calls this an introductory price. The gains come mainly from the model "working harder": it takes more reasoning steps and calls tools iteratively, so it can consume more tokens per task, especially at higher effort levels. A gated sibling, 3.8 Flash Cyber, shares the same base model. It ships with looser cyber mitigations and was tuned to prioritise vulnerability fixing over exploitation.

Evidence

General model

• HLE-Verified: 54.9%. No baseline is given.
• DeepSWE v1.1, Vals Finance Agent V2, Harvey Legal Agent Benchmark: the page claims 3.8 Flash beats 3.7 Flash and "most larger frontier models" but gives no numbers or competitor names.
• Prompt injection: the page claims a "significant leap" in robustness as measured by Gray Swan, with no figures.

Cyber variant

• CWE-Bench (Collinear, patching): pass@1 of 47.2%, against 47.8% for an unnamed "leading frontier model" at what the page calls significantly higher cost.
• Internal benchmark: vulnerability discovery across 20 programming languages, with a success rate above 70%. This benchmark was built because CyberGym covers only C/C++.
• CyberGym: the page says the model beats 3.5 Flash Cyber and larger frontier models, with no numbers.
• Chrome Security: 2.6× more correct patches to Chrome vulnerabilities than "the best commercial models", which are unnamed.
• Wiz internal pentest benchmark: 7.5–9.7% higher recall at 2.3–5.2× lower cost than other leading frontier models, which are unnamed.
• Google Cloud Vulnerability Research: found a "critical foundational vulnerability" in under 2 hours. This is a single anecdote with no details.

How to use it

3.8 Flash is available through:

• the Gemini API (Google AI Studio)
• Android Studio
• Google Antigravity
• Stitch
• Gemini Enterprise
• the Gemini app, AI Mode and Sheets, for AI Pro and Ultra subscribers

Effort levels: use lower effort levels to limit token overhead.

3.7 Flash remains "fully supported" for workloads where efficiency comes first. The page explicitly recommends it when compute cost is the main constraint.

3.8 Flash Cyber is available only through the Fairwind Program, by application. It is aimed at government authorities, critical-infrastructure operators and software maintainers.

Limits

• Real cost may rise. The per-token price is unchanged, but each task may use more tokens. Benchmark your own workloads at each effort level before assuming the upgrade is cost-neutral.
• Pricing may change. The price is labelled "introductory", and the page does not say for how long it applies.
• Comparisons can't be checked from the page. Most comparative claims have no numbers, and the competitor models are never named. The 20-language benchmark is internal, and the Chrome and Wiz results come from their own internal evaluations.
• Patching is roughly at parity, not ahead. On CWE-Bench the Cyber variant is about level with the frontier model (47.2% vs 47.8%); its advantage there is cost.
• Safeguards differ between the variants. 3.8 Flash ships with CBRN and cyber-offense safeguards under Google's Frontier Safety Framework, which may limit security tooling built on the general model. The Cyber variant relaxes these safeguards, which is why it is gated.
• Model details are not given. The page gives no model size, context length or API model identifier.