Gemini 3.8 Flash and 3.8 Flash Cyber

Sep 02, 2026 10:44 PM - 1 hour ago 1

Our newest Gemini models present next-generation intelligence for agentic workflows and cybersecurity.


Raluca Ada Popa

Gemini Security Lead, Google DeepMind


a leader image reference "Gemini 3.8 Flash and 3.8 Flash Cyber"

Building connected the momentum of 3.7 Flash from 3 weeks agone and marking our 3rd Flash merchandise successful only six weeks, coming we’re introducing Gemini 3.8, our champion reasoning & coding exemplary yet, astatine the aforesaid velocity and debased costs of 3.7. Gemini 3.8 introduces 2 variants:

  • Gemini 3.8 Flash: our astir intelligent workhorse model, delivering important improvements from 3.7 Flash crossed package engineering, agentic tasks, and critical, multi-step reasoning successful specialized domains. It is disposable astatine the aforesaid introductory value 1 arsenic 3.7 Flash astatine $0.75 per cardinal input tokens and $3.75 per cardinal output tokens.
  • Gemini 3.8 Flash Cyber: our astir tin cybersecurity exemplary pinch frontier-level capacity successful vulnerability discovery and automated patching, disposable to trusted defenders done our caller Fairwind Program.

While tailored for different deployment environments, some of today's releases are powered by the aforesaid foundational intelligence, and further accelerated by long-running agentic loops designed to recursively measure and refine the underlying models. The important coding and reasoning gains crossed this shared halfway were driven by a number of innovations, including rigorous training successful the highly demanding domain of cybersecurity.

Gemini 3.8 Flash: built for long-horizon coding and autonomous agents

Gemini 3.8 Flash delivers important gains from 3.7 Flash, often approaching the capacity of higher-cost frontier models.

comparison floor plan showing Gemini 3.8 Flash and different models

On DeepSWE v1.1 (Long-Horizon Software Engineering) 3.8 Flash outperforms astir larger frontier models successful autonomously solving analyzable engineering problems extremity to end, only astatine a fraction of the cost.

chart showing Gemini 3.8 Flash DeepSWE v1.1

Additionally, 3.8 Flash exhibits the dependability required for captious endeavor autonomy, crossed specialized knowledge domains. In quantitative and master fields that require precocious study and reporting, 3.8 Flash outperforms 3.7 Flash and different frontier models successful benchmarks for illustration Vals Finance Agent V2 and Harvey's Legal Agent Benchmark. 3.8 Flash besides achieves a 54.9% connected HLE-Verified, demonstrating its expertise to grip multi-step reasoning crossed STEM, humanities, and master fields.

chart showing Gemini 3.8 Flash Vals Finance Agent v2

a floor plan showing Gemini 3.8 Flash Harvey's Legal Agent Benchmark

a floor plan showing Gemini 3.8 Flash HLE-Verified

These capacity gains stem from a halfway creation choice: 3.8 Flash useful harder. On analyzable tasks, it exhibits greater diligence — executing other reasoning steps, and calling devices iteratively. At times, the exemplary mightiness usage much tokens to maximize performance, particularly astatine higher effort levels.

For applications wherever compute ratio is the superior constraint, developers tin utilize little effort levels to minimize token overhead aliases proceed to trust connected Gemini 3.7 Flash, which remains afloat supported for efficiency-first workloads.

Gemini 3.8 Flash Cyber: master cyber performance

Gemini 3.8 Flash Cyber, disposable to a group of trusted defenders via the Fairwind Program, provides a decisive advantage successful today’s analyzable cybersecurity landscape, pinch the Flash velocity and costs that enables speedy iteration.

Autonomous vulnerability discovery

On the modular manufacture benchmark for uncovering vulnerabilities, CyberGym, Gemini 3.8 Flash Cyber demonstrates frontier-level capacity successful autonomous vulnerability discovery. It surpasses some 3.5 Flash Cyber arsenic good arsenic importantly larger frontier models.

a floor plan showing Gemini 3.8 Flash Cyber CyberGym Pass@1

To amended seizure real-world protect needs which are not constricted to conscionable C/C++ codebases for illustration successful CyberGym, we besides evaluated Gemini 3.8 Flash Cyber against a broad soul benchmark successful which the exemplary has to observe a wide scope of vulnerabilities crossed analyzable codebases spanning 20 programming languages. Here, the exemplary showcases an awesome leap complete our erstwhile models and reaches a occurrence complaint exceeding 70%.

Chart showing Gemini 3.8 Flash Cyber existent world vulnerability discovery

Automated patching

With Gemini 3.8 Flash Cyber, we focused specifically connected equipping defenders pinch master capabilities that springiness them an advantage complete attackers. This is why we person invested successful vulnerability fixing from the start, and prioritized it complete violative capabilities for illustration exploitation.

CWE-Bench, tally by Collinear, is simply a challenging outer benchmark for patching capabilities. On this benchmark, Gemini 3.8 Flash Cyber is connected the Pareto frontier: pinch a pass@1 of 47.2% compared to a starring frontier exemplary astatine 47.8%, yet offered astatine a importantly little cost.

Chart showing Gemini 3.8 Flash Cyber StaticBench Pass@1 vs Cost Per Rollout

Real-world impact: securing Google’s code

We’re already utilizing Gemini 3.8 Flash Cyber to unafraid codification crossed Google. For example:

  • The Chrome Security squad recovered that 3.8 Flash Cyber produced 2.6 times much correct patches to vulnerabilities successful Chrome than the champion commercialized models that are overmuch larger.
  • Wiz recovered that Gemini 3.8 Flash Cyber achieves +7.5-9.7% higher callback connected their soul penetration testing benchmark for a 2.3-5.2x little costs compared to different starring frontier models.
  • Google’s Cloud Vulnerability Research squad leveraged the 3.8 Flash Cyber exemplary to find a captious foundational vulnerability successful little than 2 hours, a vulnerability for which investigation and find usually takes months.

What our Fairwind Program partners are saying

quote from David Slater, Founder and Chief Architect, Armadin

quote from Charlie Sestito, Director, Office of the CTO, Palo Alto Networks

quote from Mayank Upadhyay, CSTO, Snowflake

quote from Gal Nagli, Head of Threat Exposure, Wiz

Built pinch information successful mind

3.8 Flash ships pinch safeguards against misuse successful the domains of Chemical, Biological, Radiological, and Nuclear (CBRN) and cyber offense, while enabling beneficial usage cases, arsenic per our Frontier Safety Framework. 3.8 Flash Cyber ships pinch a much permissive group of mitigations for cybersecurity, and arsenic such, is only disposable to trusted defenders who require a much broad group of cyber capabilities.

Gemini 3.8 models person besides made a important leap successful punctual injection robustness arsenic measured by Gray Swan, protecting Gemini exemplary users from prompt-injection related malicious attacks.

a floor plan showing Gemini 3.8 Flash Cyber Gray Swan IPI Benchmark

Gemini 3.8 Flash and Cyber: get started today

  • Developers: Build pinch 3.8 Flash and research agent-first workflows successful Google Antigravity aliases commencement building coming successful the Gemini API via Google AI Studio and Android Studio, aliases make UIs successful Stitch. Get started pinch our developer docs.
  • Enterprises: Access 3.8 Flash successful Gemini Enterprise.
  • Consumers: 3.8 Flash is disposable to Google AI Pro and Ultra subscribers crossed the Gemini app, AI Mode successful Google Search and Gemini successful Google Sheets.
  • Cyber: Through our caller Fairwind Program, we’re providing trusted authorities authorities, arsenic good arsenic captious infrastructure operators and package maintainers pinch prioritized entree to Gemini 3.8 Flash Cyber. Apply for access.
More