Gemini 3.8 Flash Release with Cybersecurity Variant

Google released Gemini 3.8 Flash and 3.8 Flash Cyber. - Third Flash release in six weeks: 3.6 end July, 3.7 mid Aug, 3.8 now. - Same price as 3.7: $0.75/M input, $3.75/M output; promo until end 2026, then $1.50/$7.50. - Leading benchmarks: Vals Finance Agent v2 61.4 vs Opus 5 58.6; Harvey legal 10.0; Terminal-bench 2.1 89.4 vs 89.1; CharXiv Reasoning 86.2; LVBench 87.8 agentic/87.1 static; HLE-Verified 54.9 vs Opus 5 54.4 and GPT-5.6 Sol 54.5; BioMystery hard 56.5; LABBench2 86.2. - Lagging benchmarks: GDPVal-AA v2 Elo 1545 vs Opus 5 1824; Terminal-bench 4.0 19.1 vs Opus 5 51.8 / GPT-5.6 Sol 37.3; OSWorld-2.0 59.0 vs Opus 5 75.4; GDP.PDF 35.0 vs GPT-5.6 Sol 40.0. - DeepSWE v1.1 73.7: 0.3 below Opus 5 74.0; above GPT-5.6 Sol 72.7, Sonnet 5 53.8, 3.7 Flash 65.3. - Gains vs 3.7: BioMystery +13, DeepSWE +8.4, OSWorld +8.4, Terminal-bench 4.0 +7.9; HLE +1.3, CharXiv +1.7, Vals +2.4. - Cost efficiency: 73-74% accuracy at $1.5-$2/task; Opus 5 $13, Fable 5 $15; long SWE 1/7-1/8 cost. - Flash progression: 3.5 33%, 3.6 46%, 3.7 65%, 3.8 74% at near constant cost. - Google: 3.8 Flash works harder with more steps/tools/tokens. - Gray Swan prompt injection: 3.8 Flash 5.5%; DeepSeek V4 Pro 60.1%; Grok 4.6 51.8%. - Cyber variant only via Fairwind Program for vetted government, critical infrastructure operators, security researchers; 650+ partners. - Cyber: CyberGym vulnerability discovery 86.2%; >70% across 20 languages; CWE-Bench auto patching 47.2% vs frontier 47.8%. - Use cases: Chrome 2.6x more vulnerabilities fixed; Wiz recall 7.5%-9.7% higher, cost 2.3-5.2x lower; Google Cloud critical vuln <2 hours. - Availability: Gemini API, AI Studio, Android Studio, Antigravity, Gemini App for Pro/Ultra, Gemini Enterprise; Cyber via Fairwind. - Context: 3.5 Pro absent; Hassabis moved to DeepMind chairman/Alphabet chief scientist; Jeff Dean left; Koray Kavukcuoglu took over Gemini.
Key insights
Gemini 3.8 Flash has mixed benchmark results
In leading benchmarks, 3.8 Flash records Vals Finance Agent v2 61.4% versus Opus 5 58.6%, Harvey legal 10.0%, Terminal-bench 2.1 89.4%, CharXiv Reasoning 86.2%, LVBench 87.8% agentic and 87.1% static, HLE-Verified 54.9%, BioMysteryBench hard 56.5%, and LABBench2 86.2%. It lags in GDPVal-AA v2 Elo by about 280 points, Terminal-bench 4.0, OSWorld-2.0, and GDP.PDF. DeepSWE v1.1 is 73.7%, 0.3 points below Opus 5 and above GPT-5.6 Sol, Sonnet 5, and 3.7 Flash. Gains over 3.7 Flash are largest in BioMystery, DeepSWE, OSWorld, and Terminal-bench 4.0.
Cost efficiency is the core argument
At 73-74% accuracy, 3.8 Flash has average task cost around $1.5-2, while Opus 5 costs about $13 and Fable 5 about $15. On long software engineering tasks it achieves nearly the same pass rate at 1/7 to 1/8 the task cost. The Flash line has moved from 3.5 Flash 33% to 3.6 Flash 46%, 3.7 Flash 65%, and 3.8 Flash 74% at near-constant cost. The promotional price is $0.75 per million input tokens and $3.75 per million output tokens until end of 2026, then doubles.
Prompt injection resistance is stronger than peers
In Gray Swan prompt injection evaluation, Gemini 3.8 Flash was broken 5.5% of the time, compared with DeepSeek V4 Pro at 60.1% and Grok 4.6 at 51.8%. Google states the model works harder on complex tasks, taking more steps, calling tools more, and consuming more tokens.
Cyber variant is restricted to vetted defense users
Gemini 3.8 Flash Cyber is available only through the Fairwind Program to reviewed government agencies, critical infrastructure operators, and security researchers. Google states that 650+ partners are currently involved and positions the model for defense-oriented vulnerability discovery and patching rather than offensive exploitation. In CyberGym vulnerability discovery it scores 86.2%; in an internal evaluation across 20 programming languages it exceeds 70% vulnerability discovery; in CWE-Bench auto patching it reaches 47.2% pass@1, near the 47.8% frontier result at lower cost.
Security use cases show deployed impact
Chrome security team reported 3.8 Flash Cyber fixed 2.6 times more Chrome vulnerabilities than the best commercial LLM. Wiz internal penetration testing found recall was 7.5% to 9.7% higher while cost was 2.3 to 5.2 times lower. Google Cloud vulnerability research used it to find a critical vulnerability in under two hours, compared with months of manual research for that class.
Flash cadence continues while Pro line is absent
The launches of 3.6 Flash at end of July, 3.7 Flash in mid-August, and 3.8 Flash now are each about three weeks apart. Meanwhile Gemini 3.5 Pro, expected in June, has not appeared. DeepMind leadership changes include Demis Hassabis moving from daily management to DeepMind chairman and Alphabet chief scientist, Jeff Dean leaving, and Koray Kavukcuoglu taking over Gemini development.
3.8 Flash availability channels
Gemini 3.8 Flash is available through Gemini API, Google AI Studio, Android Studio, Antigravity, and the Gemini app for Google AI Pro and Ultra subscribers; enterprise access is through Gemini Enterprise. 3.8 Flash Cyber requires Fairwind Program application and is not available to ordinary developers.