Days after Gemini 4 Carbon went wide, Google introduced Gemini 4 Hydrogen, dropping the “Flash” name, 9to5Google reports. Google calls it a “major standalone upgrade” over Gemini 3.8 Flash. These are Google’s claims:
- 1 million tokens of input and up to 1 million of output.
- Tasks finish 2.5x faster at the default high setting, and partners report 2x to 2.8x faster runs with 12% to 23% fewer tokens.
- Terminal-Bench 4.1 goes from 18.2% to 57%, and it edges out Argon on the Artificial Analysis Intelligence Index (51 vs. 50, 54 at xhigh).
- Better agent behavior: up to 81% less overreach, 12x fewer misaligned actions and shortcuts, and half as likely to bypass permission boundaries on blocked tasks.
- In the Gemini app, it sticks to your tone and format and calls the right Connected Apps more often.
My take
The agent-behavior numbers interest me more than the benchmarks. A fast model that also stops pushing past permission walls is what you want in a bot that runs unattended. I’ll wait for independent tests before I believe the honesty score.
God Bless, and May the Source Be with You!
