One Flash for everyone, one for the vetted few

Google's announcement on September 2 introduces two models with one name. Gemini 3.8 Flash is generally available under the model ID gemini-3.8-flash. Gemini 3.8 Flash Cyber, which Google calls "Our most capable cybersecurity model with frontier-level performance in vulnerability detection and automated patching," is "available to trusted defenders through our new Fairwind Program."

The public model

The Gemini API model page describes it as "Our most intelligent Flash model, engineered for long-horizon software engineering, autonomous agents, and complex enterprise workflows." The limits it lists:

| Field | Value | |---|---| | Input tokens | 1,048,576 | | Output tokens | 65,536 | | Inputs | Text, Image, Video, Audio, and PDF | | Output | Text | | Thinking | low, medium, high (no minimal) |

Audio generation, image generation and the Live API are not supported.

The price, and its expiry date

The announcement gives "$0.75 per million input tokens and $3.75 per million output tokens," then adds a sentence that is easy to skip: "Introductory price expires on December 31, 2026. Starting January 1, 2027, $1.50/1M input tokens and $7.50/1M output tokens."

The pricing page confirms the doubling, and shows something the announcement does not say: Gemini 3.7 Flash carries the identical rows, "$0.75 through December 31, 2026. $1.50 starting January 1, 2027" on input, "$3.75 through December 31, 2026. $7.50 starting January 1, 2027" on output. Context caching follows the same pattern for both models: "$0.075 through December 31, 2026. $0.15 starting January 1, 2027," with storage at "$0.50 / 1,000,000 tokens per hour" now and "$1.00" from January. The new model, in other words, costs exactly what the old one does, and both get more expensive on the same day.

The gated model

The Cyber variant is not on the pricing page. The DeepMind model page lists its numbers: "86.2%" pass@1 on CyberGym, a "success rate exceeding 70%" on real-world vulnerability discovery across "20 programming languages," "47.2%" pass@1 on CWE-Bench, and a "6.0%" attack success rate on the Gray Swan indirect prompt injection benchmark. The announcement adds the claim that will get quoted: "The Chrome Security team found that 3.8 Flash Cyber produced 2.6 times more correct patches to vulnerabilities in Chrome than the best commercial models." None of these come with a methodology link on the pages fetched for this story.

Access runs through the Fairwind Program, which "gives high-priority defenders (like governments, healthcare providers, and telecommunications services) early access to advanced models that help them build better defenses, before new threats arrive." Priority goes to governments and national cyber authorities, critical infrastructure operators, and "Core technology platforms serving millions of users." Google says it will "vet all applicants to make sure they have a proven track record of ethical operations and research."

The conditions are operational, not contractual boilerplate. Partners must run user-level authentication with phishing-resistant MFA, restrict the model to "internal cybersecurity, incident response, or penetration testing teams only," track employee access, and not share, redistribute or sell it. Permitted work is "Authorized threat simulation, reverse engineering, and malware analysis for defensive and academic research purposes."

The companion Google blog post by Four Flynn, Vice President for Security and Privacy, pairs the model with CodeMender, Google's code-fixing agent, and puts the program at "more than 650 participating partners globally" on launch day. There is no public date for wider availability of the Cyber model.

Why the split matters

This is the third frontier lab in a month to ship a cyber-capable model behind a gate rather than a price. XT.PT has covered OpenAI's Astra and Anthropic's Mythos line. Google's version is the most explicit about the controls it expects on the customer's side, and the only one that publishes the eligibility list on a form you can fill in.

Primary sources: Introducing Gemini 3.8 Flash and 3.8 Flash Cyber, Gemini 3.8 Flash model page, Gemini API pricing, Gemini 3.8 Flash Cyber, Fairwind Program, Google's Fairwind Program blog post, read 2026-09-08.