Google has released Gemini 3.8 Flash and a second version called Gemini 3.8 Flash Cyber, announced on the Google blog.

On Hacker News, commenter mattlondon points to a leaderboard at deepswe.datacurve.ai where the new model currently sits on top, ahead of Opus 5. He also cites artificialanalysis.ai, which lists an intelligence score of 59 for Gemini 3.8 Flash, matching Opus 5 medium. He calls that surprising for a Flash model, Google's cheaper and faster tier, and says the real test is how it performs in day-to-day use.

Developer Simon Willison ran his usual pelican-drawing test at three thinking-effort settings. The high-effort run cost 8.9742 cents; his prior test on Gemini 3.7 cost 8.4387 cents. He also generated an HTML page with the new model for 1.8 cents in 13 seconds, and says its speed combined with strong HTML and JavaScript output is exciting. Separately he notes the low-effort setting looks like a regression compared with 3.7.

Not every commenter got that far. Andreygrehov says he opened gemini.google.com after the release and found only Flash-Lite, 3.6 Flash, and 3.1 Pro in the model picker, no 3.8 Flash. He says that gap is why he does not stick with Google's products: if a model is not available the moment it is announced, he moves on to whatever OpenAI or Anthropic ships next.

Commenter jampa, who uses Gemini 3.7 for a personal trip-planning app, says it beats what he tried elsewhere on real-world knowledge such as opening hours and geography, on ordering places into a sensible route, on picking the best photo of a location, and on pulling trip details out of PDFs. He tells other users not to discount Gemini if their work is not coding.

One reply asks whether Google has given up on building frontier models for outside users at all.