Google launches Gemini 3.8 Flash for AI agents

Google introduces Gemini 3.8 Flash, an AI model designed for programming, autonomous agents and tasks that require multiple reasoning steps. It arrives three weeks after Gemini 3.7 Flash and keeps the same speed and entry-level price: $0.75 per million input tokens and $3.75 per million output tokens.

The company is also announcing Gemini 3.8 Flash Cyber, a specialized variant for detecting vulnerabilities and proposing security fixes. This second version will not be available to everyone: Google is offering it to verified defenders through its new Fairwind Program.

A model that works harder when the task requires it

Gemini 3.8 Flash is designed to handle lengthy jobs, such as modifying an entire software project, analyzing legal documents or preparing financial reports. On complex tasks, it can spend more steps reasoning and use tools multiple times before responding.

That improves results, but it can also increase token usage. Developers will be able to choose lower effort levels to prioritize cost and speed, or continue using Gemini 3.7 Flash in applications where efficiency matters more than analytical depth.

According to tests shared by Google, Gemini 3.8 Flash outperforms most larger frontier models on DeepSWE v1.1, a benchmark that measures whether an AI can solve software engineering problems from start to finish. It also improves on its predecessor in financial and legal task evaluations.

On HLE-Verified, a reasoning evaluation covering science, the humanities and professional fields, it scores 54.9%. These figures come from benchmarks, so they do not guarantee the same performance in every real-world application.

A version specialized in cybersecurity

Gemini 3.8 Flash Cyber shares the general model's foundation but is trained to assist security teams. Its main goal is to find and fix vulnerabilities, not exploit vulnerable systems.

On CyberGym, a standard test for measuring automated vulnerability discovery, Google says it outperforms Gemini 3.5 Flash Cyber and larger frontier models. In an internal evaluation using code written in 20 programming languages, it achieved a success rate above 70%.

For vulnerability repair, it reached 47.2% pass@1 on CWE-Bench, a metric that measures whether the model's first proposal solves the problem. That result is close to the 47.8% achieved by a leading model, but Google says 3.8 Flash Cyber costs considerably less.

The company is already using it with its own teams. According to the published data:

• Chrome's security team achieved 2.6 times more correct patches than with the best larger commercial models.
• Wiz recorded 7.5 to 9.7 percentage points more recall, meaning a greater ability to detect existing flaws, at a cost between 2.3 and 5.2 times lower than other frontier alternatives.
• Google Cloud's vulnerability research team found a critical flaw in less than two hours. Google notes that discovering vulnerabilities of this kind usually takes months, although this case does not necessarily represent typical performance.

What changes for you

If you develop software, Gemini 3.8 Flash can handle longer tasks: reviewing several files, running tests, fixing errors and repeating the process until it completes a goal. The advantage is not just generating code, but maintaining a workflow for longer.

Businesses can access the model through Gemini Enterprise. Developers can try it in the Gemini API, Google AI Studio, Android Studio, Google Antigravity and Stitch. Google AI Pro and Ultra subscribers will also find it in the Gemini app, Google Search's AI Mode and Gemini in Google Sheets.

Gemini 3.8 Flash Cyber is restricted to government authorities, critical infrastructure operators and software maintainers accepted into Fairwind. The restriction exists because its protections for cybersecurity tasks are more permissive than those of the general model.

Google says both versions include safeguards against misuse in areas such as cyberattacks and chemical, biological, radiological or nuclear materials. It also says it has improved their resistance to prompt injection, a technique that attempts to make the model ignore its rules through malicious text.

The important part of this launch is not just that Google is releasing another Flash model within a few weeks. The company is turning fast, inexpensive models into tools that can maintain complex work for longer. The next point to watch will be whether that performance holds up outside benchmarks and how much it really costs to leave these agents working autonomously.