Google casts Gemini 3.8 Flash as a faster, stronger workhorse for coding and AI agents, while outside assessments see a more complicated bargain: frontier-level performance may come with a larger token bill. Its restricted cyber variant adds a second tension—powerful defensive automation, but only for vetted users.

Google rolled out Gemini 3.8 Flash as its third Flash update in six weeks, saying the model makes major gains in software engineering, agentic work and multi-step reasoning. Sundar Pichai said it outperforms most larger frontier models on DeepSWE v1.1, a software-engineering benchmark.Sundar Pichai (sundarpichai)¹ The company’s pitch is that Flash can now take on harder jobs without surrendering its speed-oriented identity.

The model is available to developers and enterprise users at an introductory $0.75 per million input tokens and $3.75 per million output tokens—the same rates as Gemini 3.7 Flash. But Google’s own caveat undercuts the simplest interpretation of that price tag: “the model might use more tokens to maximize performance, especially at higher effort levels.”² Developers focused on controlling consumption can remain on 3.7 Flash.

That distinction matters because 3.8 Flash is designed to “work harder,” using additional reasoning steps and iterative tool calls on complex assignments. Artificial Analysis estimated that its overall cost at this intelligence level was about 40% higher than 3.7 Flash, driven by more output tokens and more agentic turns, despite unchanged per-token pricing.²

Benchmark results offer Google a potent counterargument. Reporting on the launch found 3.8 Flash leading DeepSWE at the discounted rate, though its improvement in agentic computer use still left it well behind Claude Opus on OSWorld-2.0.³ In other words, Google is claiming a broad leap; independent framing suggests a particularly meaningful advance in coding, not a clean sweep everywhere.

Alongside the mainstream model, Google introduced Gemini 3.8 Flash Cyber for governments and trusted Fairwind Program partners. Pichai called it the company’s “most capable cybersecurity model,” built to find vulnerabilities and patch them at scale.Sundar Pichai (sundarpichai)⁴ Google says the system includes safeguards against CBRN and cyber-offense misuse, while positioning it as a tool to protect critical infrastructure.²