By admin 
Google continues its rapid-fire release schedule of advanced AI models, unveiling two powerful new iterations of its Gemini Flash family: Gemini 3.8 Flash and Gemini 3.8 Flash Cyber. This latest announcement, made on Wednesday, showcases Google’s commitment to pushing the boundaries of AI performance and accessibility, particularly in complex reasoning tasks and the critical domain of cybersecurity. The company is positioning these new models as significant leaps forward, offering enhanced capabilities at competitive price points and broadening the applications of its generative AI technology.
Gemini 3.8 Flash, described as a "workhorse" model, is engineered to excel in agentic tasks, software development, and intricate multi-step reasoning. This version builds upon the foundation of its predecessor, Gemini 3.7 Flash, delivering what Google CEO Sundar Pichai described on X as "significant leaps" across these key areas. Notably, 3.8 Flash has demonstrated superior performance on the DeepSWE coding benchmark, outperforming many larger, more resource-intensive frontier models while operating at a substantially lower cost. This cost-effectiveness, coupled with its advanced reasoning abilities, makes it a compelling option for developers and businesses seeking to leverage AI for complex problem-solving and code generation.
Complementing the general-purpose 3.8 Flash is Gemini 3.8 Flash Cyber, a specialized variant meticulously optimized for vulnerability detection and mitigation. Pichai highlighted this model as Google’s "most capable" cybersecurity AI to date. Flash Cyber is designed to match frontier-level performance in identifying security weaknesses and automating their remediation. Its effectiveness is underscored by impressive benchmark scores: 86.2% on the CyberGym cybersecurity benchmark and 47.2% on CWE-Bench, a metric that specifically evaluates AI’s ability to patch vulnerabilities. Furthermore, an internal Google benchmark revealed that Flash Cyber achieved a success rate exceeding 70% in discovering vulnerabilities across an impressive 20 different programming languages.
This marks Google’s third Flash model release in a mere six weeks, demonstrating an unprecedented pace of innovation and iteration in the AI space. The swift succession of versions, particularly following closely on the heels of 3.7 Flash, signals Google’s agile development methodology and its drive to rapidly deploy cutting-edge AI solutions to the market.
3.8 Flash: Working "Harder" with "Greater Diligence"
Gemini 3.8 Flash is immediately accessible within Gemini Enterprise, offering developers a powerful new tool. It can also be explored and implemented through the Gemini API, accessible via Google AI Studio, Google Antigravity, and Android Studio. Additionally, it supports UI generation within Stitch, further integrating it into developer workflows. The introductory pricing for 3.8 Flash mirrors that of 3.7 Flash, set at $0.75 per million input tokens and $3.75 per million output tokens. A key feature of the Gemini Flash models, including 3.8, is the user’s ability to customize and adjust "model effort levels." This allows for a fine-tuned balance between quality, cost, and latency, catering to diverse project requirements. For instance, users prioritizing compute efficiency can opt for lower token overhead, or, if maximal efficiency is the sole concern, they can continue to utilize 3.7 Flash, which Google senior product director Tulsee Doshi and Gemini security lead Raluca Ada Popa confirmed remains "fully supported for efficiency-first workloads" in their official blog post.
Doshi and Popa elaborated on the enhanced capabilities of 3.8 Flash, stating, "3.8 Flash works harder," exhibiting "greater diligence" when tackling complex tasks. This diligence translates to the model’s willingness to execute additional reasoning steps to achieve optimal performance, even if it means consuming more tokens. The model boasts a substantial 1 million token input window, capable of processing vast amounts of information, and a 64,000 token output limit. Its multimodal capabilities extend beyond text to include images, audio, video, and PDF files, enabling a more comprehensive understanding and generation of diverse content.
The performance of 3.8 Flash has been rigorously evaluated across a wide spectrum of benchmarks. These assessments cover critical areas such as coding proficiency, multimodal understanding, computer usage, long-context comprehension, knowledge work, and scientific reasoning. Google asserts that the model also demonstrates exceptional aptitude in specialized knowledge domains that demand in-depth analysis and comprehensive reporting. For example, on benchmarks like Vals Finance Agent V2 for financial analysis and Harvey’s Legal Agent Benchmark for legal applications, 3.8 Flash not only surpassed its predecessor but also outperformed other leading frontier models. Its prowess in multi-step reasoning tasks across disciplines like mathematics, science, and humanities was further validated by a score of 54.9% on Humanity’s Last Exam (HLE)-Verified.
To illustrate the practical applications of 3.8 Flash, Google shared several compelling examples. In one instance, using Google’s Antigravity platform, the model was prompted to build a game. The resulting creation was a sophisticated 3D experience featuring puzzles, dynamically changing storytelling influenced by the in-game environment, and integrated imagery and textures from Nano Banana, allowing a wizard to navigate a castle. In another demonstration, 3.8 Flash developed a fully functional DOS version of Google Maps, complete with interactive locations, turn-by-turn directions, and street views, showcasing its ability to reconstruct complex systems with a retro aesthetic. The model also generated a 3D visualizer that automatically decomposed devices into layers for detailed inspection, featuring a slider for interactive exploration. Furthermore, it created a detailed topographic map of renowned geographical sites, drawing on real datasets from the U.S. Geological Survey. This map included dynamic cross-sections, 2D projections, and scientifically accurate explanations, highlighting the model’s capacity for data interpretation and visualization.
The impressive capabilities of 3.8 Flash are further substantiated by its rankings on Arena.ai. The model debuted at No. 14 in Agent Arena, surpassing DeepSeek-V4-Pro and marking a significant improvement over Gemini 3.7 Flash, which ranked much lower at No. 32. In Text Arena, it secured the No. 7 position, outperforming notable models like Claude Opus 5 and its predecessor, Gemini 3.7 Flash. Its performance gains over 3.7 Flash were evident across multiple categories, including multi-turn requests, writing, literature and language tasks, handling longer queries, complex prompts, coding, instruction following, software and IT services, and business, management, and financial operations.
Flash Cyber: Already Securing Google’s Code
Gemini 3.8 Flash Cyber is initially being rolled out to a select group of "trusted defenders" through Google’s Fairwind Program. This initiative prioritizes government authorities, critical infrastructure operators, and other strategic partners who require advanced cyber defense capabilities. Interested organizations can apply for access to this cutting-edge technology.
Google emphasizes that Flash Cyber has undergone "rigorous training" specifically within the cybersecurity domain, representing a "significant leap in prompt injection robustness." Its primary strengths lie in autonomous vulnerability discovery and automated patching, according to internal Gemini benchmarks. Raluca Ada Popa also highlighted its strong coding abilities in a video presentation. The overarching objective behind Flash Cyber’s development is to equip cybersecurity defenders with expert-level capabilities, providing them with a crucial advantage over threat actors, whether they be malicious AI agents or human hackers. Doshi and Popa clarified their development philosophy: "We have invested in vulnerability fixing from the start, and prioritized it over offensive capabilities like exploitation."
Due to its more permissive set of mitigations for cybersecurity safeguards, which is essential for its advanced offensive detection capabilities, Flash Cyber is currently restricted to limited partners. This controlled release ensures that the model’s powerful features are not inadvertently misused. Safeguards are in place to prevent its application in cyber offense and in sensitive areas such as chemical, biological, radiological, and nuclear (CBRN) threats.
Google is already leveraging 3.8 Flash Cyber internally to enhance the security of its own vast codebase. In tests involving Chrome vulnerabilities, the model produced 2.6 times more correct patches compared to much larger commercial security models, demonstrating its superior efficacy. Wiz, a cybersecurity firm acquired by Google earlier this year for a record $32 billion, reported that 3.8 Flash Cyber exhibited a 7.5% to 9.7% higher recall of real-world vulnerabilities on an internal penetration testing benchmark. This was achieved at a significantly lower cost, between 2.3 to 5.2 times less than leading frontier models. Similarly, Google’s Cloud Vulnerability Research team identified a critical foundational vulnerability in less than two hours using 3.8 Flash Cyber, a task that typically takes months of dedicated research and discovery, according to Google’s claims.
Popa underscored the growing threat landscape, noting that AI agents are "incredibly skilled" at both finding and exploiting vulnerabilities. The challenge for defenders is immense: scanning vast codebases with large AI models is prohibitively expensive, and security teams are often overwhelmed. "In cybersecurity, attackers need only find one significant flaw over millions of lines of code. Defenders have to remove every one of those flaws to be able to defend against attackers," she explained, highlighting the asymmetrical nature of cyber warfare.
Doug Turner, engineering director for Chrome, described a recent "vulnerability apocalypse" driven by the proliferation of generative AI. "Simply overnight, we saw a hockey stick increase in the number of software vulnerabilities reported through our vulnerability research program," he stated in a video. He provided a striking example of a vulnerability discovered by 3.8 Flash Cyber that had persisted in Chromium and Chrome for an astonishing 13 years. This "very subtle bug" had eluded the attention of dozens, if not hundreds, of engineers who had reviewed the code without flagging it. Turner concluded optimistically, "Gemini 3.8 Flash Cyber is going to allow us to create better suggested fixes so that developers’ lives can get a lot easier." This sentiment encapsulates the transformative potential of these new AI models in not only identifying but also efficiently rectifying the most complex and persistent security flaws, thereby bolstering the digital defenses of organizations worldwide.
