Google Unveils Two New Gemini Flash Models Simultaneously

Google announced two new models in its "Gemini" AI series: "Gemini 3.8 Flash" and "Gemini 3.8 Flash Cyber." The former is designed primarily for AI agents (AI that autonomously performs multiple tasks on behalf of humans) and complex reasoning tasks, while the latter specializes in vulnerability detection and remediation in the cybersecurity field.

Gemini 3.8 Flash represents Google's third Flash model deployed within six weeks, arriving immediately after its predecessor, 3.7 Flash. Google's CEO stated that the model achieved significant improvements over the previous version in coding, agent processing, and multi-step reasoning domains. For instance, on the DeepSWE benchmark, which measures coding performance, it surpassed many large-scale cutting-edge models while achieving results at lower cost.

Pricing is set at $0.75 per 1 million input tokens and $3.75 per 1 million output tokens, matching the pricing at the time of Gemini 3.7 Flash's launch. The context window (the amount of information processable at once) supports up to 1 million input tokens and 64,000 output tokens, and accepts images, audio, video, and PDFs as input in addition to text. Users can adjust the model's processing volume according to their priorities for quality, cost, and processing speed. According to Google's blog, users who prioritize computational efficiency may continue using Gemini 3.7 Flash, with both models intended for selection based on use case requirements.

In terms of performance, it surpassed predecessor models and competing advanced models on "Vals Finance Agent V2," which evaluates financial agents, and "Harvey's Legal Agent Benchmark," which assesses legal tasks. Additionally, on the test version of "Humanity's Last Exam (HLE)," a collection of challenging problems spanning diverse academic fields, it achieved a score of 54.9%. Google positions this as a capability to handle problems requiring multi-step reasoning across mathematics, science, and humanities.

Regarding the cybersecurity-specialized version, Flash Cyber, Google's CEO stated it is "the most capable of the company's cybersecurity models." In vulnerability detection and large-scale patch application, it reached performance levels comparable to state-of-the-art standards, achieving 86.2% on the cybersecurity benchmark "CyberGym" and 47.2% on "CWE-Bench," which evaluates AI patch application capabilities. Additionally, Google's internal benchmarks showed over 70% success rates in vulnerability detection across 20 programming languages.

Gemini 3.8 Flash is already available for Gemini Enterprise users, and developers can test it through the Gemini API via Google AI Studio, Google Antigravity, Android Studio, and the UI design tool Stitch. As AI agent practical applications advance, the approach of separating general-purpose reasoning capabilities and cybersecurity as a specialized domain into distinct models reflects an enterprise implementation style focused on selecting the optimal model for specific use cases.