Google 三周内再推 Gemini 3.8 Flash：介绍价 $0.75/$3.75，Cyber 网络安全变体齐发Google 三周内再推 Gemini 3.8 Flash：介绍价 $0.75/$3.75，Cyber 网络安全变体齐发

On 2026-09-02 Google shipped two models at once: Gemini 3.8 Flash and Gemini 3.8 Flash Cyber — the third Flash release in six weeks, only three weeks after 3.7 Flash. 3.8 Flash is billed as the 'most intelligent workhorse', generally available in Google AI Studio and the Gemini API under the ID gemini-3.8-flash, ready for production.

On capability, 3.8 Flash scores 73.7% on the long-horizon software engineering benchmark DeepSWE v1.1 (3.7 Flash: 65.3%), approaching the much larger Claude Opus 5 (74.0%); 54.9% on HLE-Verified; and beats Opus 5 and GPT-5.6 on financial-agent (Vals Finance Agent V2, 61.4%) and Harvey legal-agent benchmarks. The design philosophy is that it 'works harder': extra reasoning steps, iterative tool calls and self-verification on complex tasks. 1M-token context, 64K output, and text/image/video/audio/PDF inputs.

Pricing matches 3.7 Flash's introductory rate: $0.75 per million input tokens and $3.75 per million output tokens through Dec 31, 2026; standard pricing doubles to $1.50/$7.50 from Jan 1, 2027. Cached input is $0.075/M (cache storage $0.50/M·hour); Batch and Flex are half price; Priority is $1.35/$6.75. Thinking tokens bill as output — press tests found ~30% more output tokens per task and roughly 40% higher real-world cost. Use thinking_level=low, or stay on the fully supported 3.7 Flash, when cost matters most.

Usage and migration: three thinking levels (low / medium default / high); the numeric thinking_budget is deprecated in favor of the thinking_level string, and the migration checklist also drops temperature/top_p/top_k and candidate_count. Available via AI Studio, Android Studio, Antigravity (its managed agent now defaults to 3.8 Flash) and Gemini Enterprise; consumers get it in the Gemini app, Search AI Mode and Sheets with AI Pro/Ultra.

The Cyber variant targets vulnerability detection and automated patching, offered to trusted defenders — governments, critical infrastructure, maintainers — through the new Fairwind Program. It reaches frontier level on CyberGym, exceeds 70% success finding vulnerabilities across ~20 languages, and posts 47.2% pass@1 on CWE-Bench patching (best commercial frontier model: 47.8% — a Pareto frontier at far lower cost). Chrome's security team saw 2.6× more correct patches than the best commercial alternative; Wiz reported 7.5–9.7% higher recall on its pentest benchmark at 2.3–5.2× lower cost.

Takeaway for developers and site owners: 3.8 Flash trades a flat sticker price for costlier tasks — more reasoning in exchange for higher task completion. Artificial Analysis pegs it near $0.58 per agentic task, among the cheapest in its performance tier. Teams doing agents, coding and knowledge work should A/B it against 3.7 at medium effort; long-running production projects should budget for the doubled 2027 price.