Google Launches Gemini 3.8 Flash and Flash Cyber Models for Advanced Coding and Cybersecurity

New Gemini Models Aim for Faster, Smarter Workflows

Google announced on September 2 2026 that it is releasing two new Gemini models designed for high‑performance coding and cybersecurity tasks.

The company described the launch as “Our newest Gemini models deliver next‑generation intelligence for agentic workflows and cybersecurity.”

Senior Director of Product Management Tulsee Doshi and Gemini Security Lead Raluca Ada Popa highlighted the rapid release cadence, noting that the announcement follows the 3.7 Flash release just three weeks earlier.

They quoted the launch statement: “Building on the momentum of 3.7 Flash from three weeks ago and marking our third Flash release in only six weeks, today we’re introducing Gemini 3.8, our best reasoning & coding model yet, at the same speed and low cost of 3.7.”

The Gemini 3.8 family consists of two variants: Gemini 3.8 Flash, a general‑purpose reasoning and coding model, and Gemini 3.8 Flash Cyber, a specialized cybersecurity model.

Both variants share a common foundational intelligence that is further accelerated by long‑running agentic loops designed to recursively evaluate and refine their outputs.

Model Improvements and Performance Benchmarks

Gemini 3.8 Flash is positioned as a “most intelligent workhorse” that delivers measurable gains over its predecessor, Gemini 3.7 Flash.

On the DeepSWE v1.1 benchmark for long‑horizon software engineering, 3.8 Flash outperforms most larger frontier models while maintaining a fraction of the cost.

In the Vals Finance Agent V2 benchmark, the model surpasses 3.7 Flash and other leading models in quantitative finance analysis.

Similarly, on Harvey’s Legal Agent Benchmark, Gemini 3.8 Flash achieves higher scores than its earlier version, indicating stronger professional‑level reasoning.

The model attains a 54.9 % score on the HLE‑Verified benchmark, demonstrating its capacity for multi‑step reasoning across STEM, humanities, and professional domains.

These performance gains are attributed to a core design choice: the model “works harder” by executing extra reasoning steps and invoking external tools iteratively.

When tackling complex tasks, the model may generate more tokens to maximize performance, especially at higher effort levels.

Developers who prioritize compute efficiency can select lower effort levels to reduce token overhead, or continue using Gemini 3.7 Flash, which remains fully supported.

Specialized Cybersecurity Offering

Gemini 3.8 Flash Cyber is targeted at “frontier‑level performance” in vulnerability detection and automated patching.

The variant is available only to trusted defenders through Google’s new Fairwind Program, ensuring controlled access to its capabilities.

Rigorous training on the demanding domain of cybersecurity underpins the model’s advanced detection and remediation abilities.

Early tests indicate that Flash Cyber can identify software vulnerabilities faster than many existing tools while also generating reliable patches.

The dedicated focus on security differentiates the Cyber variant from the general‑purpose Flash model, even though both draw on the same core intelligence.

Cost Structure and Real‑World Applications

Gemini 3.8 Flash retains the introductory pricing of its predecessor: $0.75 per million input tokens and $3.75 per million output tokens.

This pricing aims to keep the model affordable for a wide range of enterprise and developer workloads.

Google showcased several creative demos built with the model, including a puzzle‑filled wizard game generated in a single prompt using Google Antigravity.

Another demo produced a fully functional DOS‑style version of Google Maps, complete with directions and Street View, also generated through a single prompt.

Additional visualizations, such as a topographic map of famous sites built from U.S. Geological Survey data, illustrate the model’s ability to handle real‑time cross‑sections and scientific explanations.

The “Hardware Anatomy” demo demonstrates an interactive 3D visualizer that automatically decomposes devices into layered, proportionally accurate renderings using Three.js.

These examples highlight the model’s capacity to generate complex, multi‑modal content without extensive hand‑coding.

Businesses looking to adopt autonomous agents can leverage the model’s long‑horizon reasoning to build more reliable, self‑directed workflows.

For workloads where token efficiency is paramount, developers can lower the effort setting or revert to Gemini 3.7 Flash without sacrificing support.

Overall, the simultaneous release of a high‑performance general model and a security‑focused variant reflects Google’s strategy to address both productivity and protection needs in AI‑driven environments.

Why This Matters: The Gemini 3.8 releases provide enterprises with faster, lower‑cost reasoning and a dedicated, high‑accuracy cybersecurity tool, expanding the practical reach of AI in critical operations.