Google has released Gemini 3.8 Flash and Gemini 3.8 Flash Cyber. Both are based on the same underlying model but optimized differently: 3.8 Flash for programming, AI agents, and complex reasoning, Flash Cyber for finding and fixing software vulnerabilities. Both models are proprietary and not available with open model weights.
Third Flash release in six weeks
3.8 Flash is the third Flash release in just over six weeks and, counting 3.5 Flash from May, already the fourth in less than four months. The flagship model Gemini 3.5 Pro, announced back in May, still hasn't shipped. According to the Wall Street Journal, Google scrapped internal 3.5 Pro candidates and is instead planning Gemini 4 as its next flagship model.
More capability, but also more compute
The context window stays at one million tokens, same as its predecessor. The model processes text, images, audio, and video, and outputs text up to 64,000 tokens. On the software engineering test DeepSWE v1.1, Google says 3.8 Flash scores 73.7 percent, putting it practically on par with Claude Opus 5 at 74 percent. In Google's own launch comparison on Terminal-Bench 2.1 for agentic coding tasks, it improves from 85.8 to 89.4 percent, narrowly ahead of Opus 5 at 89.1 percent.
Artificial Analysis confirms the performance jump but frames it more soberly. At high reasoning effort, 3.8 Flash scores 59 points on the Intelligence Index, three more than 3.7 Flash. That puts it roughly level with GPT-5.6 Sol at its second-highest reasoning setting, but below Claude Opus 5's maximum of 63 points. According to Artificial Analysis, the progress comes mainly from better agentic performance in tool use and coding.
The API price stays at 0.75 US dollars per million input tokens and 3.75 dollars per million output tokens, same as its predecessor, though only as an introductory price: costs double starting January 2027. According to Google, 3.8 Flash applies more compute to complex tasks and can therefore consume more tokens. That shows up in Artificial Analysis's numbers too: despite identical token prices, the cost per task rises about 40 percent compared to 3.7 Flash, to 0.58 dollars. The model produces roughly 30 percent more output tokens on average, and despite around 300 output tokens per second, a task takes 2.5 minutes on average instead of 2.2. At 0.58 dollars, 3.8 Flash still remains the cheapest model tested at this performance level. For better efficiency, Google recommends lower reasoning settings or sticking with 3.7 Flash. 3.8 Flash is available through the Gemini API, Google AI Studio, and Gemini Enterprise, as well as for Google AI Pro and Ultra subscribers in the Gemini app and in Google Search's AI mode.
Flash Cyber: access through Google's new security initiative
Flash Cyber specializes in cybersecurity and carries fewer restrictive safeguards in that domain. Google therefore makes it available only to select users and says it prioritizes fixing vulnerabilities over offensive capabilities like exploit development.
On CyberGym, Google says Flash Cyber solves 86.2 percent of tasks in a single attempt, narrowly ahead of GPT-5.5-Cyber at 85.6 percent. On the patching benchmark CWE-Bench, it reaches 47.2 percent versus 47.8 percent for Fable 5, though according to Google at significantly lower cost. At the Chrome team, the model reportedly produced 2.6 times more correct patches than much larger commercial models, and Google's own security researchers reportedly used it to discover a critical vulnerability in under two hours.
Access runs through the new Fairwind Program, aimed primarily at government agencies, critical infrastructure operators, and key technology providers. Google already names more than 650 partners. Fairwind thus resembles Anthropic's Project Glasswing and OpenAI's Daybreak, which give verified defenders access to capable cyber AI.