What happened
Google released Gemini 3.8 Flash, a budget model for coding and reasoning, plus a specialized cybersecurity version called 3.8 Flash Cyber. It scores 73.7% on the DeepSWE v1.1 benchmark, just below Claude Opus 5's 74.0% and above GPT-5.6 Sol's 72.7% and the previous 3.7 Flash's 65.3%. The launch price is $0.75 per million input tokens and $3.75 per million output tokens, rising to $1.50 and $7.50 starting January 2027.
Why it matters
This is the third Flash release in six weeks, while the frontier models Gemini 3.5 Pro and Gemini 4 remain missing. Google says the new model beats Anthropic's Opus 5 and OpenAI's GPT-5.6 Sol on many benchmarks despite being far cheaper per token. Independent platform Artificial Analysis gives it an Intelligence Index score of 59, on par with GPT-5.6 Sol and Grok 4.6, but notes the cost per task has risen about 40% versus 3.7 Flash, to $0.58.
What to watch
The cybersecurity model 3.8 Flash Cyber is not publicly available—Google distributes it through the Fairwind Program to government agencies, critical infrastructure operators, and software maintainers. On the CyberGym benchmark for C/C++ vulnerabilities, it scores 86.2%, beating the previous 3.5 Flash Cyber (77.5%) and GPT-5.5-Cyber (85.6%).
Summaries like this, in your inbox every morning.
Google's release cadence for its Flash series has accelerated sharply, with three launches in six weeks. This comes as its flagship frontier models, Gemini 3.5 Pro and Gemini 4, have not yet arrived. The new Deepmind head Koray Kavukcuoglu has stated that Google still aims to lead on raw capability, suggesting the budget-line focus is not a retreat from frontier work.
The performance data shows Gemini 3.8 Flash nearly matching much more expensive models on key benchmarks. On DeepSWE v1.1, it trails Claude Opus 5 by just 0.3 points while costing a fraction per token. Independent assessments from Artificial Analysis confirm the competitive position, though they also flag a nuance: the model "works harder" on complex tasks, increasing token consumption and raising cost per task by about 40% versus 3.7 Flash.
The specialized cybersecurity release follows a controlled distribution model, limited to vetted defenders through the Fairwind Program. Its benchmark results show particular strength in automated patching, nearly matching the leading frontier model at 47.2% Pass@1 on CWE-Bench while presumably costing much less.
Pick your industry and the AI tools you use, and get news related to your work every day.
Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →
Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.
Red Hat and Nvidia expanded their partnership around Red Hat AI Factory with Nvidia, adding agent runtime cont…
Pichai said on the Big Technology Podcast that Google Cloud forgoes short-term revenue for compute trade-offs…
The Department of Homeland Security ordered civilian federal agencies to remove Deepseek products, including s…
Writing as a non-engineer, the author says a service built with AI slipped three months past its planned relea…
One developer runs multiple Grok Bots, each with a single responsibility — a master bot for shared rules, sepa…
In Claude Code 2.1.283, a PermissionRequest hook receives the AskUserQuestion tool_input, adds an answers map…
