Gemini 3.8 Flash launches as the 'workhorse' model for software engineering, autonomous agents, and deep reasoning

The defensive option: Gemini 3.8 Flash Cyber

Gemini 3.8 Flash Cyber shares the same underlying reasoning and coding capabilities as the standard model but adds specialized training for vulnerability discovery, validation, and automated patching. The general model keeps standard safeguards against chemical, biological, radiological, nuclear, and cyber misuse. The Cyber edition relaxes those cyber-specific mitigations, so it won't be offered through the public API; instead, it's available through the Fairwind Program to vetted defensive agencies, critical infrastructure operators, software maintainers, and security researchers.

Finding vulnerabilities

CyberGym tests C/C++ vulnerability discovery. For 3.8 Flash Cyber, Pass@1 is 86.2%, ahead of 3.5 Flash Cyber's 77.5% and GPT-5.5-Cyber's 85.6%. Google computed its own results; the other models' numbers come from model cards and public leaderboards, so run conditions aren't perfectly aligned.

Google also ran an internal dataset across 20 programming languages, covering 1,200 recent historical vulnerabilities. 3.8 Flash Cyber succeeded on 71.0% of them; 3.7 Flash scored 58.9%, while 3.5 Flash Cyber scored 46.6%. The broader language coverage is notable, but outside teams can't yet independently reproduce this dataset or the run.

Automated patching

The Cyber model is designed to find vulnerabilities first and then fix them. Google says it prioritized remediation ahead of offensive capabilities like exploitation from the start of development. On CWE-Bench, which tests whether a model produces correct patches, benchmark owner Collinear AI ran the evaluation with the Antigravity agent harness at high thinking effort. 3.8 Flash Cyber's Pass@1 is 47.2%, just below Fable 5's 47.8%, but its average cost per rollout is far lower.

Google says the model is already in use inside its own defensive workflows and those of partners: Chrome's security team landed 2.6 times more correct patches than with the best large commercial model; Wiz measured a 7.5–9.7 point improvement in internal penetration test recall with 2.3–5.2x lower cost; and Google Cloud's vulnerability research team found a critical infrastructure flaw in under two hours that would normally take months of research. Chrome and Cloud figures come from Google; Wiz reported its own internal test results. None of the three has been independently replicated under a unified protocol.

Access controls and prompt injection

Fairwind now covers more than 650 partner organizations. Participants undergo background and security record checks, and must enforce user identity verification, phishing-resistant MFA, access controls, and employee usage tracking. Access can't be resold, shared, or redistributed. Approved use cases include authorized threat simulation, reverse engineering, and malware analysis for defensive or academic purposes.

Gray Swan independently ran indirect prompt injection tests using its combined attack set with transfer only and no computer use. The metric is attack success rate, where lower is better. Under this specific setup, 3.8 Flash's success rate came in lower, but that doesn't mean prompt injection is solved for agents connected to browsers, email, codebases, and third-party tools.

Who should upgrade

• Stick with 3.7 Flash if tasks are short, structurally stable, and high-volume, with latency and cost as the main priorities.
• Move to 3.8 Flash if work needs long context, multi-step reasoning, sustained tool calls, or code modification, and failed runs are expensive to fix manually.
• Apply for 3.8 Flash Cyber if your organization has a legitimate need for vulnerability research or patching and can meet Fairwind's identity, permission, and audit requirements.

Where to get it

• Developers: Call 3.8 Flash through the Gemini API and Google AI Studio, or use it in Google Antigravity, Android Studio, and Stitch.
• Enterprises: The general model is available in Gemini Enterprise.
• Consumers: Google AI Pro and Ultra subscribers can access it in Gemini, AI Mode in Google Search, and Google Sheets.
• Security teams: Apply for Flash Cyber through the Fairwind Program. Approved teams can use it standalone or with CodeMender. Access via managed models on the Gemini Enterprise Agent Platform supports zero data retention. Organizations that don't qualify for Fairwind can still pair CodeMender with the public model.

Before upgrading, benchmark 3.7 and 3.8 on the same harness with identical tool permissions and acceptance criteria. Test at least 20–50 real tasks and track completion rate, total tokens, end-to-end latency, tool spend, and how often a human has to intervene. The default medium thinking setting isn't right for every workload; try low for simple tasks and high for critical ones.

The model can still hallucinate, slow down intermittently, or time out. Google lists the knowledge cutoff as March 2026, though some areas may still reflect January 2025 data. In internal automated safety evals, compared with 3.7 Flash, multilingual safety metrics slipped 5.4 points; Google says human review found most of the regression came from false positives or non-severe samples. The core case for Gemini 3.8 Flash is better completion rates on long-horizon work. Whether it's worth replacing 3.7 comes down to completion rate and total cost on your actual tasks.