What makes Gemini 3.8 Flash different from earlier Flash models, and why has Google introduced a separate Cyber version?
Google introduced Gemini 3.8 Flash and Gemini 3.8 Flash Cyber on September 2, 2026, expanding the Gemini Flash family with a general-purpose model and a specialized cybersecurity variant.
Gemini 3.8 Flash is positioned for more demanding reasoning, coding, and agentic tasks, while Gemini 3.8 Flash Cyber is designed specifically for defensive security use cases.
But the bigger questions are how much these models actually improve performance, what they cost to run, how their benchmarks compare, and where each model is available. This guide examines those areas in detail.
Gemini 3.8 Flash is Google’s most intelligent Flash model, engineered for long-horizon software engineering, autonomous agents, and complex enterprise workflows while retaining the speed and cost efficiency of the Flash family. Google introduced gemini-3.8-flash on September 2, 2026, as a generally available model ready for production use.
Gemini 3.8 Flash accepts text, images, video, audio, and PDFs as input and generates text output. It supports an input context window of 1,048,576 tokens and a maximum output of 65,536 tokens (Source).
Google provides three configurable thinking levels: low, medium, and high, with medium used by default. The minimal thinking level is not supported by Gemini 3.8 Flash. This lets developers adjust reasoning effort according to the complexity, latency, and cost requirements of a task.
The model also supports function calling, code execution, file search, Google Search grounding, Google Maps grounding, structured outputs, URL context, and Computer Use. Computer Use remains a preview capability, although Google currently identifies Gemini 3.8 Flash as its recommended model for computer-use workflows.
Gemini 3.8 Flash focuses on improving multi-step reasoning, iterative tool use, agentic execution, and complex software-engineering tasks rather than introducing a completely new set of surface-level capabilities (blog.google).
Google says Gemini 3.8 Flash can perform additional reasoning steps and call tools iteratively when a task requires more work. This improves its ability to:
These changes make the model better suited to workflows where success depends on repeated reasoning and execution rather than a single response (deepmind.google).
Gemini 3.8 Flash also improves agentic coding and long-horizon software engineering. Compared with earlier Flash capabilities, it is designed to stay engaged with more complex development tasks that require planning, code changes, tool use, validation, and iteration across multiple steps (ai.google.dev).
Gemini 3.8 Flash’s benchmark results and pricing highlight the model’s central trade-off: stronger performance on complex coding, reasoning, and agentic tasks, with costs that can rise as reasoning effort and token usage increase.
Google’s September 2026 evaluations show Gemini 3.8 Flash improving over Gemini 3.7 Flash across several coding, agentic, and expert-reasoning benchmarks. (deepmind.com)
Google’s evaluation table combines self-computed results with scores from public leaderboards and external benchmark providers, so individual benchmarks should be interpreted within their specific evaluation methodology.
The strongest gains are concentrated in software engineering and agentic execution, while improvements across finance, legal, and broader expert reasoning are more incremental.
According to Google’s official Gemini API pricing, the output rate includes thinking tokens.
This means actual task cost depends not only on the listed per-token rate but also on how much reasoning the model performs. Google notes that Gemini 3.8 Flash may consume more tokens on difficult or long-running tasks, particularly at higher thinking levels. For less demanding workloads, developers can lower the thinking level to reduce token usage and control cost. (ai.google.dev)
Gemini 3.8 Flash Cyber is Google’s specialized Gemini 3.8 model for defensive cybersecurity. Introduced alongside Gemini 3.8 Flash on September 2, 2026, it is powered by the same foundational intelligence as the standard model but is tailored for a different deployment environment (Source).
Google describes Gemini 3.8 Flash Cyber as its most capable cybersecurity model, with capabilities focused on finding and fixing software vulnerabilities. Unlike the standard Gemini 3.8 Flash, which serves broader reasoning, coding, and agentic workloads, Flash Cyber is specifically intended for advanced cybersecurity tasks.
The models also differ in their safety configuration. Gemini 3.8 Flash includes safeguards against cyber-offensive misuse, while Flash Cyber uses a more permissive set of cybersecurity mitigations to support defenders who require a broader range of cyber capabilities. For this reason, Google limits Gemini 3.8 Flash Cyber access to trusted defenders through its Fairwind Program.
In short:
Gemini 3.8 Flash Cyber is designed around two connected defensive-security tasks: discovering software vulnerabilities and helping remediate them. Google evaluates these capabilities using external benchmarks, internal multilingual codebase testing, and real-world security environments.
The distinction matters because vulnerability discovery and vulnerability fixing are different challenges. A model may identify insecure code correctly but still fail to produce a patch that resolves the issue without affecting normal software behavior.
Vulnerability discovery measures how effectively the model can examine software, reason about code behavior, and identify security weaknesses (Google DeepMind).
Google reports the following results:
Pass@1 measures whether the model successfully solves a benchmark task on its first generated attempt. A higher score therefore indicates stronger first-attempt reliability.
The two evaluations also test different things:
Google’s published CyberGym results also place Gemini 3.8 Flash Cyber ahead of Gemini 3.5 Flash Cyber and several larger frontier models included in the same evaluation.
Overall, the discovery results point to stronger performance in both benchmark-based vulnerability research and broader multilingual software analysis.
Finding a vulnerability is only the first step. A defensive-security model must also be able to generate a fix that removes the weakness while preserving the intended behavior of the software.
Google reports several results for Gemini 3.8 Flash Cyber’s remediation capabilities:
CWE-Bench is especially useful because it evaluates whether a model can successfully patch vulnerable software, not simply recognize that a vulnerability exists.
Google also reports results from operational security environments:
These results come from different evaluation environments, so they should not be treated as directly comparable scores:
Taken together, the evidence supports Google’s positioning of Gemini 3.8 Flash Cyber as an end-to-end defensive-security model capable of supporting both vulnerability discovery and remediation across benchmark and real-world environments.
Gemini 3.8 Flash and Gemini 3.8 Flash Cyber target different workloads. Gemini 3.8 Flash is the general-purpose model for reasoning, coding, tool use, and agentic workflows, while Gemini 3.8 Flash Cyber is specialized for defensive cybersecurity tasks such as vulnerability discovery and patching.
For most developers and organizations, Gemini 3.8 Flash will be the relevant option. Flash Cyber is intended for a narrower group of trusted defenders who need its more permissive cybersecurity capabilities and are approved for Google’s Fairwind Program.
The key decision is not which model has the highest benchmark score, but which one best matches your workload, reasoning requirements, expected token usage (including thinking tokens), and access constraints.
Gemini 3.8 Flash is Google’s most intelligent Flash model, designed for long-horizon software engineering, autonomous agents, complex workflows, and multi-step reasoning while retaining Flash-level speed and efficiency.
Google released Gemini 3.8 Flash on September 2, 2026. The stable gemini-3.8-flash model is generally available through the Gemini API.
Gemini 3.8 Flash Cyber is Google’s specialized cybersecurity model, built on the same foundational intelligence as Gemini 3.8 Flash but optimized for defensive tasks such as vulnerability discovery and patching.
Google introduced Gemini 3.8 Flash Cyber on September 2, 2026, alongside Gemini 3.8 Flash and the launch of its Fairwind Program for trusted defenders.
Gemini 3.8 Flash is generally available (GA). Google provides the stable gemini-3.8-flash model for production use, with no shutdown date currently announced.
Gemini 3.8 Flash Cyber is not generally available. Access is restricted to approved trusted defenders through Google DeepMind’s Fairwind Program, which applies governance and due-diligence requirements to participating organizations.
Developers can access Gemini 3.8 Flash through the Gemini API, Google AI Studio, Android Studio, Google Antigravity, and Stitch. Google also offers it through Gemini Enterprise and to eligible consumers across the Gemini app, AI Mode in Google Search, and Gemini in Google Sheets.
Gemini 3.8 Flash Cyber is available through the Fairwind Program to approved trusted partners. Eligible organizations can use it as a standalone model or together with Google DeepMind’s CodeMender for vulnerability research and patching workflows.
