Three Flash releases in six weeks, and the interesting line isn’t a benchmark — it’s Google admitting 3.8 Flash “works harder,” burning extra reasoning steps and iterative tool calls to win. Same sticker price as 3.7, but tokens are the meter, not the rate card. Anyone running long agent loops against a cheap model should read that as a variable bill wearing a fixed-price hat.
What I’d do: keep 3.7 on the efficiency-first jobs — Google says it stays fully supported — and put 3.8 behind the tasks where finishing matters more than token count, at a lower effort level until I’ve watched the spend. Then diary the introductory price: it doubles January 1, 2027. Also worth noting the claimed jump in prompt-injection robustness, which matters more than SWE scores if your agent touches anything you didn’t write.
The story — Google introduced Gemini 3.8 Flash and 3.8 Flash Cyber on September 2, 2026. Flash targets long-horizon coding and autonomous agents, scoring 54.9% on HLE-Verified at 3.7’s introductory price of $0.75/$3.75 per million tokens. Flash Cyber, for vulnerability discovery and automated patching, reaches 47.2% pass@1 on CWE-Bench and is limited to trusted defenders via the new Fairwind Program. (Source)