Bindu Reddy Says Gemini 3.8 Flash Regresses on Benchmarks

Bindu Reddy, CEO of Abacus.AI, posted that Gemini 3.8 Flash scores lower than 3.7 Flash on the company's benchmark. She wrote that the new version appears overfit to public benchmarks, performs worse on hidden questions, and regresses on data analysis. Reddy called for Google to stop weekly Flash releases and ship Gemini 4.0 instead. In a follow-up she noted frequent checkpoints leave Google trailing open-source efforts. A separate post from Da7_Tech added that 3.7 Flash handles agentic tasks more reliably while 3.8 works better only as a chat model.