Google Unveils Gemini 3.8 Flash for Agents and Cyber for

Google today announced a significant expansion of its AI model lineup, unveiling two distinct versions of its new Gemini 3.8 Flash. Released on Wednesday, September 2, 2026, the updated models include a standard Flash designed as a versatile "workhorse" for agentic tasks, software development, and complex multi-step reasoning, alongside a specialized Flash Cyber variant optimized for advanced vulnerability detection and mitigation in cybersecurity. This rapid release marks Google's third Flash model update in just six weeks, following closely on the heels of version 3.7.

Google CEO Sundar Pichai highlighted that Gemini 3.8 Flash represents "significant leaps" over its predecessor, 3.7 Flash, particularly in software engineering and agentic capabilities. It has demonstrated superior performance on the DeepSWE coding benchmark, often outperforming larger frontier models while maintaining a considerably lower cost. The model is immediately available in Gemini Enterprise and accessible to developers via the Gemini API through platforms like Google AI Studio and Android Studio.

Gemini 3.8 Flash: The Agentic Workhorse

Pricing for 3.8 Flash is set at $0.75 per million input tokens and $3.75 per million output tokens, matching the introductory rates of 3.7 Flash. Users can adjust the model's effort levels to balance quality, cost, and latency, with 3.7 Flash remaining fully supported for efficiency-first workloads. Google senior product director Tulsee Doshi noted that 3.8 Flash "works harder," exhibiting "greater diligence" in complex tasks, potentially using more tokens to maximize performance.

The model boasts a substantial 1M-token input window and a 64K-token output limit, allowing it to process diverse data formats including text, images, audio, video, and PDF files. Benchmarked across various domains such as coding, multimodal capabilities, and scientific reasoning, 3.8 Flash excelled. It outperformed previous models on specialized benchmarks like Vals Finance Agent V2 and Harvey's Legal Agent Benchmark, also achieving a 54.9% score on Humanity’s Last Exam (HLE)-Verified for multi-step reasoning.

Google showcased its versatility with examples like generating a 3D wizard game using simple prompts in Antigravity, creating a functional DOS version of Google Maps, and developing a 3D visualizer for device inspection. Independent evaluations from Arena.ai placed 3.8 Flash at No. 14 in Agent Arena, a substantial improvement from 3.7 Flash's No. 32, and No. 7 in Text Arena, surpassing notable competitors like Claude Opus 5.

Flash Cyber: Defending Against a Vulnerability Apocalypse

The Gemini 3.8 Flash Cyber model is being strategically rolled out to "trusted defenders" through Google’s Fairwind Program, which grants access to government authorities, critical-infrastructure operators, and partners seeking advanced cyber defense. This model has undergone "rigorous training" in the cybersecurity domain and exhibits significant robustness against prompt injection attacks.

Pichai confirmed Flash Cyber as Google's "most capable" cybersecurity model, matching frontier-level performance in discovering and patching vulnerabilities at scale. It scored 86.2% on the CyberGym benchmark and 47.2% on CWE-Bench for AI patching. Internally, it achieved over a 70% success rate in discovering vulnerabilities across 20 programming languages. The model prioritizes vulnerability fixing over offensive capabilities, safeguarding against misuse in cyber offense and CBRN (chemical, biological, radiological, and nuclear) contexts.

Google is already leveraging 3.8 Flash Cyber to secure its own codebase, demonstrating 2.6 times more correct patches for Chrome vulnerabilities compared to larger commercial models. Acquisitions like Wiz, valued at $32 billion, have benefited, with Flash Cyber showing 7.5% to 9.7% higher recall of real-world vulnerabilities at 2.3 to 5.2 times lower cost than leading frontier models on internal penetration testing benchmarks. Google's Cloud Vulnerability Research team also reported finding a critical foundational vulnerability in under two hours, a process that typically takes months.

Doug Turner, Engineering Director for Chrome, highlighted a recent "vulnerability apocalypse" driven by generative AI, leading to a "hockey stick increase" in reported software vulnerabilities. He noted that 3.8 Flash Cyber even uncovered a "very subtle bug" in Chromium and Chrome that had persisted for 13 years despite numerous engineers examining the code. Turner emphasized that Flash Cyber will significantly ease developers' lives by providing better suggested fixes.

FAQ

Q: What are the two main versions of Google’s Gemini 3.8 Flash models?

A: Google has released a standard Gemini 3.8 Flash, a general-purpose "workhorse" for agentic tasks, software development, and multi-step reasoning, and Gemini 3.8 Flash Cyber, which is specifically optimized for cybersecurity tasks like vulnerability detection and mitigation.

Q: How does Gemini 3.8 Flash compare to its predecessor, 3.7 Flash?

A: Gemini 3.8 Flash delivers "significant leaps" in performance across software engineering and agentic tasks, outperforming 3.7 Flash and many frontier models on benchmarks like DeepSWE and improving significantly in Arena.ai rankings. While offering enhanced diligence for complex tasks, it retains similar introductory pricing.

Q: Who can access Gemini 3.8 Flash Cyber, and what is its primary focus?

A: Gemini 3.8 Flash Cyber is initially available to "trusted defenders" through Google’s Fairwind Program, targeting government authorities, critical-infrastructure operators, and partners. Its primary focus is autonomous vulnerability discovery and automated patching, prioritizing defensive capabilities and safeguarding against misuse.