Google launches Gemini 3.8 Flash with improved reasoning but warns of higher token usage
Google's newest model claims better performance on complex tasks but may cost more to run despite unchanged per-token pricing.
What happened
Google released Gemini 3.8 Flash, claiming the model performs more reasoning steps and calls tools iteratively compared to its predecessor Gemini 3.7 Flash. The new model maintains the same introductory pricing ($0.75 per million input tokens, $3.75 per million output tokens) but Google warns it "might use more tokens to maximize performance, especially at higher effort levels." According to Artificial Analysis, Gemini 3.8 Flash is the cheapest model measured at its intelligence level, though it costs approximately 40% more than 3.7 Flash due to 30% higher output tokens per task. The model outperforms competitors including Anthropic's Fable 5 on software engineering and agent benchmarks (DeepSWE v1.1, Vals Finance Agent V2, and Harvey's Legal Agent benchmarks). Google also released Gemini 3.8 Flash Cyber through its new Fairwind Program, limited to governments and trusted partners including CrowdStrike and the Center for Internet Security, which provides access to a CodeMender agent for finding and fixing vulnerabilities.
Context
The rapid release cycle—3.8 Flash arriving weeks after 3.7 Flash—suggests Google is iterating quickly in the competitive generative AI market. The trade-off between improved reasoning and increased token usage creates a practical decision point for developers choosing between performance gains and cost predictability. The Fairwind Program's restricted access to Gemini 3.8 Flash Cyber reflects differentiated access models for sensitive cybersecurity applications. Google's emphasis on software engineering and agentic performance, combined with competitive benchmarking against Anthropic's models, indicates a focus on enterprise and developer adoption in these specific domains.