Google releases Gemini 3.8 Flash and a cybersecurity variant

Google released Gemini 3.8 Flash, a point update to the Gemini 3.7 Flash model it had shipped three weeks earlier, alongside a separate cybersecurity build, Gemini 3.8 Flash Cyber. It kept the same introductory price — $0.75 per million input tokens and $3.75 per million output, both due to roughly double on 1 January 2027 — and is available through the Gemini API, the Gemini app for paid subscribers, Google Search’s AI Mode, Sheets and Gemini Enterprise.

Google positioned the model around agentic work and reasoning, saying it spends extra reasoning steps and calls tools more persistently on hard tasks. On the company’s own comparison table — set against Gemini 3.7 Flash, Claude Opus 5, Claude Sonnet 5 and OpenAI’s GPT-5.6 Sol and Terra — Gemini 3.8 Flash led several rows despite being far cheaper than the frontier models it was measured against: 73.7% on DeepSWE v1.1, 89.4% on Terminal-Bench 2.1, and 54.9% on the HLE-Verified reasoning set, though the larger Claude Opus 5 stayed ahead on longer-horizon agent tests such as Terminal-Bench 4.0 and OSWorld-2.0. All the figures are Google’s, at benchmark versions of its choosing, rather than independently verified.

The more distinctive release was Gemini 3.8 Flash Cyber, which Google described as offering frontier-level vulnerability detection and automated patching, reporting an internal success rate above 70% at discovering vulnerabilities across 20 programming languages. Rather than shipping it openly, Google restricted the model to vetted defenders through an application-gated “Fairwind” programme — the same containment logic other labs have applied to dual-use cyber capabilities, keeping an offensive-capable tool inside a trusted-access wall while still putting it to defensive use.