Unveiling Gemini 3.8 Flash and 3.8 Flash Cyber

Date: September 02, 2026 Authors: Tulsee Doshi (Senior Director, Product Management) & Raluca Ada Popa (Gemini Security Lead, Google DeepMind)

Google is pushing the boundaries of agentic workflows and cybersecurity with the launch of its latest AI models. Following the momentum of the 3.7 Flash release just three weeks ago—and marking the second third Flash iteration in a mere six-week window—we are proud to introduce Gemini 3.8. This version represents our most advanced reasoning and coding model to date, while maintaining the high velocity and affordability of its predecessor.

---

🚀 The Dual-Variant Approach

Gemini 3.8 arrives in two distinct flavors, both sharing a foundational intelligence core but optimized for different operational environments.

1. Gemini 3.8 Flash

The "intelligent workhorse" of the family. It provides a massive leap over 3.7 Flash in several key areas:

• Software Engineering: Enhanced capabilities for complex builds.
• Agentic Tasks: Better autonomy in executing multi-step goals.
• Specialized Reasoning: Superior performance in critical, domain-specific logic.

2. Gemini 3.8 Flash Cyber

A specialized powerhouse designed for the front lines of digital defense. This model offers frontier-level proficiency in:

• Vulnerability Detection: Identifying security flaws with high precision.
• Automated Patching: Generating fixes for discovered vulnerabilities.

Access Note: Gemini 3.8 Flash Cyber is not open to the general public; it is available exclusively to trusted security professionals through the newly established Fairwind Program.

---

🛠️ Technical Architecture & Innovation

Both models are built upon the same core intelligence, but their performance is amplified by long-running agentic loops. These loops allow the models to recursively analyze, evaluate, and polish their own outputs.

Interestingly, the rigorous training required for the cybersecurity variant actually improved the general reasoning and coding capabilities of the standard 3.8 Flash model.

The "Work Harder" Philosophy

The secret to the performance gain is that Gemini 3.8 Flash is designed to be more diligent. When faced with a complex problem, it doesn't just guess; it:

1. Executes additional internal reasoning steps.
2. Calls tools iteratively until the solution is verified.

Note on Efficiency: Because the model "works harder," it may consume more tokens to achieve higher accuracy.

• For Max Performance: Use higher effort levels.
• For Max Efficiency: Use lower effort levels or stick with Gemini 3.7 Flash.

---

📊 Performance & Benchmarks

Gemini 3.8 Flash often rivals the performance of much larger, more expensive frontier models.

Key Metrics

• DeepSWE v1.1: Outperforms most large-scale models in autonomous, end-to-end software engineering.
• Professional Domains: Superior analysis in finance (Vals Finance Agent V2) and law (Harvey's Legal Agent Benchmark).
• General Intelligence: Achieved a score of $\text{HLE-Verified} = 54.9\%$, proving its versatility across STEM and the humanities.

Comparison Data

---

💻 Real-World Application

To demonstrate its autonomy, Gemini 3.8 Flash created a fully functional game within Google Antigravity. The model utilized a looping instruction to build complex puzzles and environmental mechanics from a simple prompt.

Example Logic Flow:

while game_not_complete:
    analyze_current_puzzle_state()
    generate_environmental_asset()
    verify_puzzle_solvability()
    if not verified:
        refine_logic()

Visual Evidence of Success

• Engineering:
• Finance:
• Legal:
• STEM/Humanities:

---

🛡️ Cybersecurity Deep Dive (Flash Cyber)

The Cyber variant is specifically tuned for the "cat-and-mouse" game of security.

• CyberGym Pass@1: High success rates in simulated security environments.
• Real-World Discovery: Proven ability to find actual vulnerabilities in production code.
• Cost-Efficiency: Maintains a superior ratio of Pass@1 vs. Cost Per Rollout on StaticBench.

---

✅ Summary Checklist

• Gemini 3.8 Flash: Released for general agentic and coding use.
• Gemini 3.8 Flash Cyber: Released for trusted defenders (Fairwind Program).
• Pricing: Maintained at $0.75 /$3.75 per million tokens.
• Core Tech: Recursive agentic loops for higher diligence.
• Performance: Frontier-level results on DeepSWE, HLE, and Legal/Finance benchmarks.