Google has introduced Gemini 3.8, launching 2 distinct models focused on automated coding and defensive security. The standard 3.8 Flash edition matches the low pricing of previous releases at $0.75 per million input tokens while beating rival frontier models in complex reasoning tests. But general developers will not get access to the cyber security edition, which is locked behind a specialized vetting program.
Google built 3.8 Flash to act as a workhorse model for software engineering tasks. On the DeepSWE v1.1 benchmark for long horizon development problems, the system solved complex programming challenges end to end at a lower operational cost than larger frontier competitors. The model also hit 54.9% on the HLE Verified test, showing a clear jump in handling multi step logic across science, mathematics, and legal analysis.
The performance gains come down to a simple technical change. The model works harder. On difficult requests, 3.8 Flash spends extra reasoning cycles and calls external tools repeatedly before spitting out an answer. That means it might burn more tokens at high effort levels, but developers who want pure speed can still adjust the settings or fall back on 3.7 Flash for simpler tasks.
Alongside the general release, Google announced Gemini 3.8 Flash Cyber. This model is engineered specifically to find software bugs and write automated fixes before attackers can exploit them. On the CyberGym benchmark, the model beat both 3.5 Flash Cyber and larger commercial systems. Internal testing across codebases spanning 20 programming languages showed a vulnerability detection success rate higher than 70%.
Automated code patching was prioritized over offensive tools during development. On the CWE Bench evaluation run by Collinear, 3.8 Flash Cyber scored a 47.2% single pass success rate, landing on par with the 47.8% score of top tier frontier systems while running at a fraction of the cost.
Google is already running the security model on internal systems. The Chrome Security team reported that 3.8 Flash Cyber created 2.6 times more correct vulnerability patches for the Chrome browser than larger commercial alternatives. Security firm Wiz saw up to 9.7% higher recall on internal penetration tests, while Google Cloud researchers used the model to isolate a foundational system flaw in under 2 hours.
General access to the Cyber edition is restricted. Google is limiting distribution to verified infrastructure operators, software maintainers, and government agencies through its Fairwind Program. For general safety, both models include mitigations against dangerous chemical or radiological misuse, alongside improved resistance against malicious prompt injection attacks based on Gray Swan evaluations.
