The Gemini 3.8 series expands on the 3.7 Flash release from three weeks ago with two new models: Gemini 3.8 Flash and Gemini 3.8 Flash Cyber. Both share the same foundational intelligence and are further accelerated by long‑running agent loops that recursively evaluate and refine outputs.
Gemini 3.8 Flash is optimized for long‑horizon coding and autonomous agent tasks. It approaches the performance of higher‑cost frontier models, outperforming 3.7 Flash on benchmarks such as DeepSWE v1.1, Vals Finance Agent V2, and Harvey's Legal Agent, and achieving a 54.9% score on HLE‑Verified. On complex problems the model adds extra reasoning steps and iteratively calls tools, which may increase token usage; developers can switch to lower‑effort modes or fall back to 3.7 Flash when compute efficiency is paramount. The model can also generate full‑featured demos in Google Antigravity from a single prompt, including a 3D puzzle game, a DOS‑style Google Maps, topographic maps built from USGS data, and an interactive hardware teardown visualizer.
Gemini 3.8 Flash Cyber targets trusted defenders via the Fairwind Program, delivering frontier‑level vulnerability discovery and automated patching at Flash speed and cost. It leads the CyberGym benchmark for autonomous vulnerability detection and, on an internal suite covering 20 programming languages, exceeds a 70% success rate. On the external CWE‑Bench patching benchmark it records a pass@1 of 47.2%, essentially matching the best frontier model while costing far less.
In production, Chrome’s security team observed 2.6× more correct patches than larger commercial models, Wiz reported a 7.5‑9.7% recall boost with 2.3‑5.2× lower cost, and Google Cloud’s Vulnerability Research team uncovered a critical flaw in under two hours—a task that normally takes months.
Safety mitigations are built into 3.8 Flash for chemical, biological, radiological, and nuclear (CBRN) domains as well as cyber offense, while 3.8 Flash Cyber offers a broader set of cyber safeguards but is restricted to vetted defenders. Prompt‑injection robustness has also been improved, as measured by the Gray Swan test.
Developers can start building with 3.8 Flash in Google AI Studio, Android Studio, or Stitch; enterprises can access it via Gemini Enterprise, and consumers can use it through the Gemini app, Google Search AI mode, and Gemini in Google Sheets. Access to Flash Cyber is provided through the Fairwind Program for government agencies, critical infrastructure operators, and software maintainers. Apply for access now.
Review