Fairwind now lists over 650 partners and offers Gemini 4 Argon as its flagship cyber model, so new applicants should expect Argon first and treat this model as the earlier, Flash-speed option. It shares the long context window and output limit of mainline Gemini 3.8 Flash, according to Google's API documentation.
What changed
Google DeepMind tuned its Gemini 3.8 Flash model for security work and announced the result on September 2, 2026. Google reports Chrome Security got 2.6 times as many correct patches as far larger commercial models. Access is limited to vetted defenders in the Fairwind Program, and the program's headline model is now the newer Gemini 4 Argon.
Provider: Google DeepMind · Family: Gemini 3.8
More about Google DeepMind on HokAI
Context window: 1,048,576 tokens · Max output: 65,536
Input modalities: text, code · Output: text, code
About Gemini 3.8 Flash Cyber
Gemini 3.8 Flash Cyber is Google DeepMind's cybersecurity fine-tune of Gemini 3.8 Flash. Google announced both on September 2, 2026, and describes them as "powered by the same foundational intelligence." It follows Gemini 3.5 Flash Cyber, and its job is narrow: find, validate and patch security vulnerabilities in real codebases, not general chat or reasoning.
Google's launch post reports frontier-level results on CyberGym, the standard benchmark for autonomous vulnerability discovery, and a success rate above 70% (71.0% in Google's later comparison chart) on an internal benchmark that spans 20 programming languages. On CWE-Bench, run by Collinear, it scores 47.2% pass@1 on patching, against 47.8% for a leading frontier model that costs significantly more per Google. The Gemini 3.8 Flash documentation lists a 1,048,576-token input limit and a 65,536-token output limit, which this variant shares.
Google cites three outside results. Chrome Security found 2.6 times more correct patches than the best commercial models that are much larger. Wiz measured 7.5 to 9.7% higher recall on its internal penetration-testing benchmark at 2.3 to 5.2 times lower cost than other leading frontier models. Google Cloud's Vulnerability Research team found a critical foundational vulnerability in under two hours, work that usually takes months.
There is no public API and no per-token price. Access runs through the Fairwind Program, an application process for government authorities, critical infrastructure operators and software maintainers. Google's program page now says it works with over 650 partners. Approved organizations must use user-level authentication and phishing-resistant MFA, may only hand the model to in-house defenders, meaning security, response and pentest staff, and may not share, resell or redistribute access. Google also runs background checks on applicants. The model ships with a more permissive set of cyber mitigations than mainline Gemini 3.8 Flash, which is the stated reason it stays gated.
The newer option is Gemini 4 Argon, announced September 30, 2026 and rolling out to Fairwind partners. Google's own chart puts Argon at 85.8% on real-world vulnerability discovery against 71.0% for 3.8 Flash Cyber, and at 70.9% against 58.2% on the Wiz penetration-test benchmark. Argon ties for first on CWE-bench v1 at 68%, a newer version of the benchmark than the one behind the 47.2% figure, so the two scores are not comparable. Google's Cyber model page now leads with Argon, and its Fairwind page describes exclusive access to Argon. Google has not said 3.8 Flash Cyber is retired, so this page stays live as a record of the earlier model.
Who should skip it: any team without Fairwind approval, anyone who needs a general coding assistant, and anyone who wants the strongest Google cyber model, which is now Argon. For another cyber-tuned model in the catalog, see GPT-5.6 Cyber. Google has not published a data retention or training policy specific to this model.
Pricing
Google has not published per-token pricing for this model. There is no self-serve API; access requires a reviewed application through the Fairwind Program at deepmind.google/fairwind-program, limited to vetted government agencies, infrastructure operators, and software-maintenance organizations.
Key Features
- CyberGym-Leading Vulnerability Discovery: Leads Google's disclosed industry benchmark for autonomous vulnerability discovery, ahead of both its predecessor and frontier models larger than itself.
- Cross-Language Vulnerability Coverage: Extends discovery across a broad, Google-disclosed set of programming languages in a single internal benchmark, the widest cross-language scope published for the Cyber lineage.
- Validated Patch Generation: Scores 47.2% Pass@1 on CWE-Bench for producing patches that resolve the underlying weakness rather than just flagging it.
- 1M-Token Context Inherited From Base 3.8 Flash: Carries the full 1,048,576-token window over from the shared Gemini 3.8 Flash base, room for a complete monorepo snapshot or a deep commit-history scan without truncation.
- Fairwind Program Gated Access: Available only to vetted applicants across government, critical-infrastructure, and software-maintenance roles who apply through Google's Fairwind Program; a wave of partner organizations were onboarded at launch.
Pros
- Real-world results come from three separate teams (Google's Chrome Security group, Wiz, and Google Cloud's Vulnerability Research unit), not only Google's internal synthetic benchmarks.
- Wiz reported higher recall on its own internal pentest benchmark, at a lower cost than other leading frontier models, which is a result from a company outside Google's model team.
- Patch quality is graded on whether a fix resolves the underlying weakness (CWE-Bench pass@1), not just whether a bug gets flagged.
- Google tested discovery across many programming languages rather than only C and C++, a wider scope than its earlier JavaScript-focused Cyber results.
Cons
- No public pricing or self-serve API; every credential needs a reviewed Fairwind Program application, which rules it out for teams that need access today.
- Google has not disclosed general-purpose benchmark scores (GPQA, MMLU-Pro, SWE-bench Verified) for this fine-tune, so buyers cannot compare it with mainline frontier models outside cyber evals.
- Looser cyber-offense safeguards than the mainline Flash model are the stated reason Google keeps it away from the public.
- Gemini 4 Argon, announced at the end of September 2026, scores higher on Google's own discovery charts, and Fairwind now centres on it.
Benchmarks
- CyberGym pass@1: 86.2 vendor-reported · 02 Oct 2026
- CWE Bench pass@1: 47.2 vendor-reported · 02 Oct 2026
- Gray Swan IPI Attack Success Rate: 6 vendor-reported · 02 Oct 2026
- Real World Vuln Discovery 20 Languages (%): 71 vendor-reported · 02 Oct 2026
A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.
Frequently Asked Questions
Is there a per-token price for Gemini 3.8 Flash Cyber?
No. Google has not published per-token pricing and there is no self-serve API or AI Studio quota for this variant. Mainline Gemini 3.8 Flash has an introductory $0.75 input and $3.75 output rate per million tokens until December 31, 2026, then $1.50 and $7.50 from January 1, 2027, but that does not apply to the gated Cyber model.
How does Gemini 3.8 Flash Cyber compare with Gemini 4 Argon?
Argon is the newer model and Google's own charts favour it: 85.8% against 71.0% for real-world vulnerability discovery. On Wiz's black-box pentest test the gap is 58.2% for 3.8 Flash Cyber versus 70.9% for Argon. Argon ties for first at 68% on CWE-bench v1, a newer version than the earlier benchmark behind this model's patching score, so those patching numbers are not comparable.
Are the weights of Gemini 3.8 Flash Cyber available to download?
No. It is proprietary and closed-weight. Unlike mainline Gemini 3.8 Flash, it is not offered through the standard Gemini API; the only route to it is a reviewed Fairwind Program application.
What does Google say about the data you send to this model?
Google's pages do not say how this model handles retained or training data. The Fairwind Program's published terms require user-level authentication and phishing-resistant MFA, limit use to internal security, incident response or penetration-testing teams, and bar partners from sharing or reselling access. Those terms are written around Argon, so check them with Google when you apply.
Who qualifies for Fairwind access, and who should skip it?
Google prioritizes government authorities, critical infrastructure operators in areas such as healthcare, telecommunications, energy and finance, and software maintainers, and it runs background checks on every applicant. Teams outside those groups, or anyone needing a general coding model, should use mainline Gemini 3.8 Flash instead.
HokAI guides covering Gemini 3.8 Flash Cyber
- Gemini 4 Argon vs GPT-6 Astra vs Claude Opus 5.5: Cost Per Real Task: Argon, GPT-6 Astra and Claude Opus 5.5 compared on what a finished task costs, not the token price, plus the 1 December price rise on Argon. Updated Oct 2026.
- Project Glasswing Explained: What Claude Mythos Found Inside Epic, and Who Else Gets Access: What Project Glasswing is, who is in it, what Epic found with Claude Mythos, and how Google Fairwind and OpenAI Codex Security compare on access and price.