KorvaJoin the waitlist
Tech

Google Does Not Use Its Smartest AI for Every Search

A faster model handles routine searches, and dials its effort up or down to save time and computing power.

AI image
TechCrunch · Google's latest Search upgrade promises better answers to your most complicated questions | TechCrunch

Google says its most powerful AI is not always needed for Search. The company says "sometimes a faster, cheaper, and reasonably smart one will do."

That model is Gemini 4 Hydrogen, which is rolling out to Search users globally. Carbon and Argon remain for the hardest reasoning work, while Hydrogen is made to work quickly.

Google describes agentic workflows as "tasks in which an AI carries out multiple steps on its own." Hydrogen is built for that kind of fast, multi-step work. In Search, that can mean breaking a question apart, running several lookups, and then assembling an answer.

The key is that Hydrogen does not always think equally hard. It has four effort settings: low, medium, high as the standard choice, plus a new xhigh option for extra effort.

Think of those levels as a dial for effort. A lower setting produces an answer faster and with fewer output tokens, the chunks of words and symbols an AI generates and that partly determine computing cost. A higher setting lets the model work longer on a harder task.

Hydrogen can also handle routine coding work on its own, such as writing software, running checks, or confirming edits. For tougher jobs, Google says, it can team up with Carbon and Argon in mixed-model jobs.

Google says this approach keeps answers fast while handling more complicated questions.

Questions people ask

What does an AI thinking level actually do?

A thinking level controls how long a model works before answering. Low means a quick, direct response, while higher levels allow more intermediate steps, checking, and refinement for harder questions.

What are output tokens, and why do they matter?

Output tokens are the pieces of text, numbers, and code a model produces. Companies pay for computing power based partly on how many tokens are generated, so fewer tokens generally means a faster and cheaper answer.

Why not use the most powerful model for everything?

The most capable models are slower and cost much more to run, especially at Google's scale of billions of searches. Using a smaller, faster model for routine work saves time and money, while reserving the heavyweight models for problems that truly need them.

More in Tech