Google on September 2 launched Gemini 3.8 Flash, a new lower-cost general-purpose model for its consumer, developer and enterprise AI products, alongside a separate cybersecurity version whose access is deliberately restricted. The split makes the release more than a routine model update: Google is putting its mainstream workhorse into broad circulation while reserving a variant with looser cyber safeguards for selected defenders.
Gemini 3.8 Flash is available through the Gemini API, Google AI Studio, Android Studio and Gemini Enterprise, as well as to Google AI Pro and Ultra subscribers in the Gemini app, AI Mode in Search and Gemini in Sheets, according to Google’s launch announcement. Its sibling, Gemini 3.8 Flash Cyber, is being offered through the Fairwind Program with prioritized access for trusted government authorities, critical-infrastructure operators and software maintainers. That eligibility list is broader than accounts describing the model as a government-only product, including early coverage by The Next Web.
A broad Flash model with long-context specifications
Flash is Google’s designation for models meant to balance capability, latency and cost in high-volume deployments. The company calls 3.8 its most intelligent workhorse model and says it improves on Gemini 3.7 Flash in software engineering, agentic work and multistep reasoning. Those are Google’s product claims, rather than independently established measurements, but the technical documentation provides useful limits for organizations deciding whether to test it.
Google’s developer guide lists Gemini 3.8 Flash as generally available. It accepts text, image, audio and video as inputs, generates text, has a context window of 1,048,576 tokens and supports outputs up to 65,536 tokens. A million-token context is designed to let an application supply a large body of material in one request—such as a code repository, long document collection or extended interaction history—though it does not by itself establish that the model will reason accurately over every part of that material.
Google kept the introductory API rate it used for 3.7 Flash: $0.75 per million input tokens and $3.75 per million output tokens. The input-output distinction can become material for coding agents and document-processing systems that generate lengthy responses; at the stated list price, output tokens cost five times as much as input tokens. Google has confirmed the introductory price, but not a later price schedule in its launch material.
The documentation also reports benchmark gains over 3.7 Flash: 90.8% versus 81.6% on Terminal-bench 2.1, 61.6% versus 60.4% on SWE-Bench Pro, and 51.9% versus 48.0% on SWE-Atlas. Such evaluations are useful directional indicators for developers, especially because they identify named tests and a predecessor comparison. They should not be read as a universal ranking. Benchmark setup, tool access, prompting and the mix of real-world software tasks can all affect how an agent performs after deployment.
The cyber model changes the availability calculation
Google says the two 3.8 models share foundational intelligence but are tailored for different settings. The cyber version is aimed at vulnerability discovery and patching, and Google explicitly says it has more permissive cybersecurity mitigations than the generally released model. Rather than making that configuration public, it is using Fairwind to screen and prioritize access among organizations it characterizes as trusted defenders.
That is a practical acknowledgment of dual-use risk. Tools that can inspect code and identify weaknesses can help a maintainer fix an exposed library or help an infrastructure operator identify a flaw before attackers exploit it. The same general capability can be misused if distributed without constraints. Google has not said that all eligible organizations automatically receive access; its wording describes a prioritized program, not an open product tier.
Google reported that 3.8 Flash Cyber exceeded a 70% success rate on an internal vulnerability-discovery benchmark spanning codebases in 20 programming languages. It also reported a 47.2% pass@1 score on CWE-Bench, compared with 47.8% for an unnamed leading frontier model. “Pass@1” means the outcome from a model’s first attempt, rather than its best result across repeated tries. The comparison offers some external benchmark context, but the unnamed rival and Google’s description of relative cost make it difficult to independently assess the full trade-off.
The company further said Chrome’s security team obtained 2.6 times more correct vulnerability patches from the model than from the best much-larger commercial models, and that Wiz observed 7.5% to 9.7% higher recall in an internal penetration-testing benchmark at lower cost. Both are potentially important deployment anecdotes, but they remain results reported by Google from internal or partner testing. Neither substitutes for a published methodology, independently repeatable evaluation or a broad measure of performance in live defensive environments.
Google’s rapid Flash release cycle
The launch follows an unusually fast cadence for this product line. Google says 3.8 Flash is its third Flash release in six weeks and arrives three weeks after Gemini 3.7 Flash. The sequence began with the July 21 release of Gemini 3.6 Flash, Gemini 3.5 Flash Cyber and Gemini 3.5 Flash-Lite, a rollout reported at the time by CNBC. Google is therefore asking customers to absorb successive model changes at roughly the same time it is widening its deployment channels.
For developers, that cadence can bring measurable model improvements without an immediate increase in the advertised entry price. It also creates work: applications need regression testing for tool use, output formatting, safety behavior, latency and costs before a model swap. For security teams, the more consequential fact is not simply that Google has built a cyber-focused Flash variant, but that it has drawn an operational boundary around who may use the version with more permissive mitigations.
