Google DeepMind has just introduced Gemini 3.8 Flash and Gemini 3.8 Flash Cyber, the two latest variants in the Flash lineup, marking the third Flash release in just six weeks. Both run on the same core intelligence foundation, accelerated by agentic loops that automatically evaluate and fine-tune the model.
Gemini 3.8 Flash is the most capable general-purpose model in the Flash family, showing significant improvements over 3.7 Flash in software engineering, agentic tasks, and multi-step reasoning across specialized domains.
Pricing remains at 3.7 Flash’s introductory rates: $0.75 per million input tokens and $3.75 per million output tokens (expiring December 31, 2026, after which it will be $1.50/$7.50).
Regarding published performance:
-
On DeepSWE v1.1 (long-horizon software engineering), 3.8 Flash surpasses most larger frontier models at a fraction of the cost.
-
On Vals Finance Agent V2 and Harvey’s Legal Agent Benchmark, it outperforms both 3.7 Flash and other frontier models.
-
HLE-Verified (interdisciplinary scientific reasoning) reached 54.9%.
In terms of real-world performance, according to Arena.ai, Gemini 3.8 Flash (High effort) is rated to outperform several popular current models:
-
Agent Arena: Reached rank 14 with a +5.94% net improvement (compared to Gemini 3.7 Flash with only +0.84% improvement) and surpassed DeepSeek-V4-Pro (#15, +5.91%).
-
Text Arena: Reached rank 7 with 1,494 points, placing ahead of Claude Opus 5 (High) (rank 8, 1,492 points), with improvements across Writing, Multi-Turn, Hard Prompts English, Coding, and Instruction Following.
-
Code Arena WebDev: Ranked 18th with 1,567 points, edging out Claude Opus 4.8 at rank 19 with 1,563 points.
The defining distinction of Gemini 3.8 Flash is its design for continuous, long-duration execution; on complex tasks, the model performs additional reasoning steps and repeated tool calls. If computational efficiency is the priority, developers can select a lower effort level or continue using 3.7 Flash.
Gemini 3.8 Flash Cyber is dedicated exclusively to security professionals, accessible only through the Fairwind program for governments, critical infrastructure operators, and software maintainers. Highlights include:
- CyberGym (automated vulnerability discovery): Outperforms both 3.5 Flash Cyber and larger frontier models.
-
Internal 20-programming-language benchmark: Vulnerability discovery success rate exceeds 70%, a sharp jump over the previous generation.
-
CWE-Bench (automated patching): 47.2% pass@1, on par with the leading frontier model (47.8%) but at significantly lower cost.
- Real-world impact at Google: Chrome Security recorded 2.6 times the number of correct patches compared to a larger commercial model; Wiz recorded 7.5–9.7% higher recall at 2.3–5.2 times lower cost; the Cloud Vulnerability Research team discovered a critical vulnerability in under 2 hours (a process that typically takes months).
Both models show substantial improvements in prompt injection resistance (according to Gray Swan benchmarks) and adhere to the Frontier Safety Framework across sensitive domains (CBRN, cyberattacks). 3.8 Flash Cyber features more permissive mitigations for cybersecurity work, which is why it is restricted to trusted users.