Google Introduces Gemini 3.8 Flash; Cyber Version Limited to Partners

On September 2, Google introduced Gemini 3.8 Flash, a generally available model intended for production use through the Gemini API. At the same time, it introduced Gemini 3.8 Flash Cyber, a separate variant for cybersecurity that is not intended for the public. Only Fairwind program partners can access it following controlled approval, and they may use it for defensive or research purposes.

The new Flash model expands Gemini’s faster, lower-cost line for programming and agent workflows. For developers, its technical limits, pricing schedule, and the way it may consume more tokens for demanding tasks are particularly important.

Gemini 3.8 Flash: One Million Tokens of Context and Three Reasoning Levels

Gemini 3.8 Flash supports a context window of one million tokens and can generate up to 64,000 output tokens. The model offers low, medium, and high reasoning levels.

Google says that for complex tasks, the model may perform multiple reasoning steps and repeatedly use tools. This is particularly relevant for AI agents that combine work with code, documentation, or external tools. However, it does not automatically mean lower costs for a specific task: more steps and tool calls may increase total token consumption.

The introductory price is set through December 31, 2026, at $0.75 per million input tokens and $3.75 per million output tokens. From January 1, 2027, the price is expected to rise to $1.50 per million input tokens and $7.50 per million output tokens.

When evaluating operating expenses, the token rate alone may therefore not be enough. Actual consumption for a completed task will also be important, especially if an application uses a higher reasoning level or tool calls.

Flash Cyber Remains in a Controlled Program

Gemini 3.8 Flash Cyber is not publicly available. Only Fairwind program partners gain access following controlled approval, exclusively for defensive and research-oriented cybersecurity use.

Google claims that this variant offers better capabilities for finding vulnerabilities and preparing fixes. However, this is not confirmation of a specific publicly known incident or proof that the model will be equally effective in all real-world security situations. The available materials also do not specify when or under what broader conditions access might be extended to additional organizations.

The restricted access model distinguishes Flash Cyber from the general model. Google is therefore placing more advanced capabilities related to finding and fixing security flaws in a controlled partner program rather than making them part of its publicly available offering.

Benchmarks Should Still Be Read Carefully

Google compares the new models using the DeepSWE, CyberGym, and CWE-Bench benchmarks, as well as its own internal tests. However, these results are primarily manufacturer claims. Based on the available materials, it is unclear to what extent the methodologies have been fully independently verified and whether the results can be reproduced by external teams.

The coming months will show independent comparisons of performance and costs per completed task. It will also be important to monitor the expansion of the Fairwind program and the specific conditions under which Google might make Gemini 3.8 Flash Cyber available to additional organizations.

Verified and updated: 09/03/2026 06:20