Google released Gemini 4 Hydrogen on Thursday, positioning it as the successor to Flash and focusing on speed when writing code and running agent tasks.
Several sets of official figures: Under the default high inference setting, the speed of completing tasks is 2.5 times that of Gemini 3.8 Flash; the Terminal-Bench 4.1 score increased from 18.2% to 57%; the number of fabricated facts was significantly reduced. Google also said that it surpassed its flagship Argon in the Artificial Analysis comprehensive index by default.
The context is an input of 1 million tokens and a maximum output of 1 million tokens. API prices are $1.50 per million inputs and $7.50 per million outputs.
The price is interesting. It is twice as expensive as 3.8 Flash, so it is called the "successor of Flash", but it is actually not in the same price range. But it’s cheaper than Argon: Argon’s current discount price is $2 / $10, and it will double starting December 1st. Google obviously wants to move a large number of people who originally used Argon for daily coding to Hydrogen.
Free users can also encounter it. The automatic model selection function in the Gemini app will switch to Hydrogen for the appropriate problem.
Also online is Orchestrator: a main agent (Carbon for Pro users, Argon for Ultra users) responsible for disassembly tasks, and then dispatched to multiple Hydrogen sub-agents to run. It has been pushed to Antigravity. Expensive models think and cheap models work. This division of labor is being tried by every company. Google has made it the default product form this time.
Benchmarks should be discounted. The third party measured about 54% on Terminal-Bench 4.0, which is not as high as the official said. In addition, some reviews found that its hallucination rate is still about twice that of Argon. It’s fast, medium-priced, and suitable for large-volume jobs. The key judgment is that it’s better to leave it to a more expensive model.