Google introduced two new artificial intelligence models this week. The models are called Gemini 3.8 Flash and Gemini 3.8 Flash Cyber.
The announcement came on September 2, 2026. It marks the third Flash model release from Google in only six weeks.
Gemini 3.8 builds on the earlier Gemini 3.7 Flash model. Google says it is their best reasoning and coding model so far, while keeping the same speed and cost as the previous version.
There are two versions of the new release. One is built for general use, and the other is built specifically for cybersecurity work.
Gemini 3.8 Flash is priced at $0.75 per million input tokens and $3.75 per million output tokens. This matches the introductory price of Gemini 3.7 Flash.
Both new models share the same underlying technology. Google says this shared core was improved through training in cybersecurity, which also helped boost general coding and reasoning skills.
How Gemini 3.8 Flash performs
Gemini 3.8 Flash is designed for long coding tasks and for AI agents that work on their own. Google says the model shows strong gains over the earlier 3.7 Flash version.
On a coding test called DeepSWE v1.1, the model outperformed most larger and more costly AI models. It did this while solving engineering problems from start to finish on its own.
The model also performed well in specialized fields like finance and law. It scored higher than 3.7 Flash and other top models on the Vals Finance Agent V2 and Harvey's Legal Agent Benchmark tests.
On a benchmark called HLE-Verified, Gemini 3.8 Flash scored 54.9%. This test measures reasoning across science, humanities, and professional subjects.
Google explained that the model performs extra reasoning steps and uses tools more than once when working on hard problems. This can mean it uses more computing power, especially at higher settings.
Developers who want to save on computing costs can choose lower effort settings. They can also keep using Gemini 3.7 Flash, which Google says will remain supported.
What Gemini 3.8 Flash Cyber does
Gemini 3.8 Flash Cyber is built for cybersecurity defense. It is only available to select trusted organizations through a new program called Fairwind.
On a benchmark called CyberGym, the model showed strong results in finding software vulnerabilities on its own. Google says it outperformed both its own earlier Cyber model and larger competing models.
Google also tested the model on an internal benchmark covering 20 programming languages. The model found a wide range of vulnerabilities and reached a success rate above 70%.
Google said it focused on building the model's ability to fix vulnerabilities rather than exploit them. On a patching test called CWE-Bench, the model scored a 47.2% pass rate, close to a leading rival model's 47.8%, but at a lower cost.
Google is already using the Cyber model inside the company. The Chrome Security team found it produced 2.6 times more correct patches for Chrome vulnerabilities than larger commercial models.
Wiz, a cloud security firm, reported the model achieved higher detection rates on penetration tests while costing several times less to run. Google's own Cloud Vulnerability Research team used the model to find a critical vulnerability in under two hours, a process that normally takes months.
