Google released Gemini 3.8 Flash and Gemini 3.8 Flash Cyber on Sept. 2, 2026, calling the pair its third Flash model drop in six weeks. Ars Technica reported the launch the same day.
Gemini API docs list model code gemini-3.8-flash, an input limit of 1,048,576 tokens, and an output limit of 65,536 tokens, with text, image, video, audio, and PDF input. Google says introductory API pricing matches 3.7 Flash at $0.75 per million input tokens and $3.75 per million output tokens through Dec. 31, 2026, then $1.50 and $7.50. The DeepMind model card sets a March 2026 knowledge cutoff for some domains, with others limited to January 2025. Google AI Pro and Ultra subscribers get 3.8 Flash in the Gemini app, AI Mode in Search, and Gemini in Sheets, while developers can use Google AI Studio, the Gemini API, Android Studio, and Google Antigravity.
Google says 3.8 Flash is its most intelligent Flash workhorse for long-horizon software engineering and agent workflows, and that it often uses more reasoning steps and tool calls on hard tasks. Company tables put the model at 73.7% on DeepSWE v1.1, 54.9% on HLE-Verified, 61.4% on Vals Finance Agent v2, 89.4% on Terminal-Bench 2.1, and 59.0% on OSWorld-2.0. Google also says 3.7 Flash remains fully supported for efficiency-first workloads.
Gemini 3.8 Flash Cyber shares that foundational stack, Google says, but ships with more permissive cyber mitigations and is limited to trusted defenders through the Fairwind Program for governments, critical infrastructure operators, and software maintainers. Company materials claim frontier-level CyberGym vulnerability discovery without stating a numeric CyberGym score in the launch post, a Collinear CWE-Bench pass@1 of 47.2%, more than 70% success on an internal 20-language vulnerability suite, and a Chrome Security finding of 2.6 times more correct patches than larger commercial models. Google says it prioritized patching over exploitation, and that ordinary 3.8 Flash keeps tighter CBRN and cyber-offense safeguards under its Frontier Safety Framework.
Independent: Artificial Analysis's live Gemini 3.8 Flash (high) page scores the model 41 on Intelligence Index v4.3 and about 278 output tokens per second at the listed $0.75 and $3.75 rates, while noting high verbosity of about 170 million Index output tokens. Ars Technica said coding gains look larger than other benches on Google's charts and that OSWorld computer-use still trails market leaders such as Claude Opus on the same company table.
What remains open is how CyberGym, CWE-Bench, and Chrome patch gains hold under outside harnesses, and how widely Fairwind access expands beyond the trusted-defender gate. Google's launch tables and partner quotes should be weighed separately from third-party Index scores until those points are settled.