Key Takeaways
- Gemini 3.8 Flash uses longer agentic loops and iterative tool calls to beat larger rivals on DeepSWE v1.1 and Vals Finance Agent V2 at lower cost.
- Gemini 3.8 Flash Cyber helps trusted defenders find and patch vulnerabilities, hitting 47.2% pass@1 on CWE-Bench and earning 2.6x more correct patches for Chrome Security.
- With $0.75 per million input tokens, the release signals a shift toward spending extra inference tokens on reliability for multi-step agentic and security workflows.
Table of Contents
- Third Flash Release in Six Weeks Targets Coding and Cyber
- Inside the Benchmark Jump: Why Working Harder Beats Working Bigger
- Cyber Variant for Defenders Moves Beyond Vulnerability Discovery
- The Skimaki Signal: Market Pressure and the Race to Close the Coding Gap
- A New Baseline for Agentic and Security Workflows
Third Flash Release in Six Weeks Targets Coding and Cyber
Google DeepMind has shipped its third Gemini Flash release in six weeks: Gemini 3.8 Flash and a security-focused sibling, Gemini 3.8 Flash Cyber.
The base model arrives with substantially stronger long-horizon coding and agentic performance, while the Cyber variant targets vulnerability discovery and automated patching for approved defenders.
Pricing for Gemini 3.8 Flash starts at $0.75 per million input tokens and $3.75 per million output tokens, matching the introductory rate of Gemini 3.7 Flash.
Inside the Benchmark Jump: Why Working Harder Beats Working Bigger
According to Google DeepMind’s official announcement, the most consequential change is not raw speed but persistent effort.
Gemini 3.8 Flash uses longer agentic loops and iterative tool calls to push performance on complex tasks, sometimes consuming more tokens to maximize accuracy.
Developers can lower effort to reduce token overhead, while Gemini 3.7 Flash remains supported for efficiency-first workloads.
- DeepSWE v1.1 — outperforms most larger frontier models in end-to-end software engineering while running at a fraction of the cost.
- Vals Finance Agent V2 — surpasses Gemini 3.7 Flash and other frontier systems in financial agent evaluation.
- Harvey’s Legal Agent Benchmark — stronger multi-step legal analysis and reporting.
- HLE-Verified — reaches 54.9%, showing multi-step reasoning across STEM, humanities, and professional domains.
The shared core was trained in part on demanding cybersecurity tasks, which sharpened the coding and reasoning gains across both variants.
Cyber Variant for Defenders Moves Beyond Vulnerability Discovery
Gemini 3.8 Flash Cyber is not a general release.
It enters a new Fairwind Program that prioritizes access for trusted government authorities, critical infrastructure operators, and software maintainers.
On the CyberGym external benchmark, the model outperforms Gemini 3.5 Flash Cyber and significantly larger frontier models in autonomous vulnerability discovery.
Against an internal benchmark spanning 20 programming languages, it clears a success rate above 70% for discovering vulnerabilities across complex codebases.
For patch generation, CWE-Bench run by Collinear places it at a pass@1 of 47.2%, close to a leading frontier model at 47.8% but at a significantly lower cost.
Early deployment data shows real-world impact: the Chrome Security team saw 2.6 times more correct patches, Wiz measured 7.5 to 9.7 percentage points higher recall at 2.3 to 5.2 times lower cost, and Google’s Cloud Vulnerability Research team found a critical foundational vulnerability in under two hours.
The model ships with safeguards from the Frontier Safety Framework and improved prompt injection robustness measured by Gray Swan.
The Skimaki Signal: Market Pressure and the Race to Close the Coding Gap
The Wall Street Journal reported ahead of the launch that Google was preparing a model known internally as ‘Skimaki’ to close an AI coding gap.
That framing matters because the benchmark profile now supports a narrower, targeted release rather than a broad generational upgrade.
A TipRanks summary noted that Gemini 3.8 Flash had already drawn favorable internal coding evaluations from Google engineers.
That anticipation lifted Google shares roughly 0.7% in after-hours trading.
For AI practitioners, the more significant signal is the shift toward models that exchange extra inference tokens for more reliable multi-step autonomy.
Gemini 3.8 Flash still trails a leading frontier model on CWE-Bench by 0.6 percentage points, but delivers that performance at a cost profile designed to make continuous patch generation practical.
A New Baseline for Agentic and Security Workflows
The release resets expectations for what a mid-sized, low-cost model can do in autonomous coding and cyber defense. For teams building AI-assisted SEO and content systems that must adapt to rapid model releases, programmatic SEO AI automation is how Andres SEO Expert approaches it — contact us.
Frequently Asked Questions
What is Gemini 3.8 Flash and when was it released?
Google DeepMind has shipped Gemini 3.8 Flash, its third Gemini Flash release in six weeks. It is a mid-sized, low-cost model focused on coding and agentic performance, and includes a security-focused sibling, Gemini 3.8 Flash Cyber.
How does Gemini 3.8 Flash improve upon Gemini 3.7 Flash?
Gemini 3.8 Flash uses longer agentic loops and iterative tool calls to push performance on complex tasks, sometimes consuming more tokens for accuracy. It outperforms Gemini 3.7 Flash on benchmarks like DeepSWE v1.1, Vals Finance Agent V2, and Harvey’s Legal Agent Benchmark.
What is Gemini 3.8 Flash Cyber and who can access it?
It is a security-focused model that targets vulnerability discovery and automated patching for approved defenders. Access is through the Fairwind Program, prioritizing government authorities, critical infrastructure operators, and software maintainers.
What are the benchmark results for Gemini 3.8 Flash Cyber?
On CyberGym it outperforms Gemini 3.5 Flash Cyber and larger frontier models. It clears above 70% success on an internal benchmark for vulnerability discovery across 20 languages. On CWE-Bench, it reaches 47.2% pass@1 for patch generation vs 47.8% for a leading frontier model.
What is the significance of Skimaki in this release?
Skimaki was an internal model name that the Wall Street Journal reported Google was preparing to close an AI coding gap. This release reflects that narrower, targeted coding focus rather than a broad generational upgrade.
Why does Gemini 3.8 Flash work harder rather than bigger?
The model exchanges extra inference tokens for more reliable multi-step autonomy, allowing it to handle complex agentic tasks better while keeping a lower cost profile. Developers can lower effort to reduce token overhead.
