Google has released two related Gemini models: a general-purpose Gemini 3.8 Flash for coding and agentic work, and Gemini 3.8 Flash Cyber, a more tightly controlled version for cybersecurity defense. The announcement, dated September 2, positions the pair as a faster and lower-cost continuation of the company’s Flash line rather than a single model with one access policy.
One foundation, two deployment paths
Google describes Gemini 3.8 Flash as its workhorse for long-horizon software engineering, autonomous agents, and complex reasoning. It says the model is available through the Gemini API, Google AI Studio, Google Antigravity, Android Studio, Gemini Enterprise, and consumer Gemini products. The introductory price is $0.75 per million input tokens and $3.75 per million output tokens, with a higher price scheduled after December 31, 2026.
The cyber version follows a different distribution model. Google says Gemini 3.8 Flash Cyber is available to trusted defenders through the Fairwind Program, with access intended for government authorities, critical-infrastructure operators, and software maintainers. That gate reflects a practical distinction: vulnerability discovery and automated patching can help defenders, but the same capabilities could be repurposed for offensive work.
The performance story is mostly a company-reported story
Google says Gemini 3.8 Flash often approaches more expensive frontier models on long-horizon software engineering and reports a 54.9% result on HLE-Verified. For the cyber model, the company reports frontier-level performance on its CyberGym evaluation, a success rate above 70% on an internal benchmark spanning 20 programming languages, and a 47.2% pass@1 result on the external CWE-Bench patching benchmark.
Those numbers are useful signals about what Google is optimizing for, but they should be read with the normal limits of vendor benchmarks. The announcement supplies the evaluation names, charts, and comparison claims; it does not provide an independent replication in the release itself. The claims therefore establish Google’s reported positioning, not a neutral league table.
Why the split matters
The more important product decision is the separation between broad availability and capability-gated access. Google is making the general Flash model easy to try while reserving the more security-sensitive variant for organizations that can demonstrate a defensive need. That approach could become a template for distributing powerful specialized models: broad access where misuse risk is manageable, and a controlled program where the same capability has obvious dual-use value.
For developers, the release increases pressure on the economics of agentic software. If Google’s performance claims hold outside its own tests, a relatively inexpensive Flash model can take on longer tasks that previously required larger systems. The questions to watch are whether the advertised price survives real workloads, how often the model calls tools, and whether the cyber program produces measurable defensive outcomes rather than benchmark gains alone.