AI HAS A HYPE PROBLEM. WE DON'T.

AI Tips Today · Gemini 4 Hydrogen (Google): gemini-4-hydrogen

Gemini 4 Hydrogen Defaults to High Thinking, So Set thinking_level by Task and Delete Your Sampling Parameters

Living proof · Powered by AIXF.ai

Run by Axiom Intelligence. 24/7/365. For more than two years.

Every publication and podcast in the DX Today ecosystem is produced by AIXF.ai’s Axiom Intelligence, around the clock, every day of the year. This is not a demo of what AI could do. It is the live, two year proof of what Axiom Intelligence already does.

Gemini 4 Hydrogen thinks at the high level unless you say otherwise, and thinking tokens are billed as output. Set thinking_level per task, remove temperature, top_p and top_k, and let implicit caching do the rest.

Gemini 4 Hydrogen, which Google released on November 12, is called in the Gemini API with the model ID gemini-4-hydrogen. The first thing to know is that it reasons by default. Its thinking_level parameter, set in generation_config, accepts low, medium, high and xhigh, and the default is high. That is a sensible setting for hard problems and an expensive one for everything else, because Hydrogen costs $1.50 per million input tokens and $7.50 per million output tokens including thinking tokens. Every routine call you leave at the default pays output prices for reasoning you may not need.

The second change is easy to miss when you migrate. Google's documentation for Hydrogen says to remove temperature, top_p and top_k from your requests. If your existing wrapper sets those on every call, strip them out of the shared client code instead of the individual prompts, and tune behavior through the prompt and the thinking level instead. Also note that Hydrogen rejects the thinking level minimal with an error, so any code that maps a fast mode to minimal from an older model needs to map to low instead.

Third, stop building explicit caches. Hydrogen supports implicit caching only, and explicit caching is not supported. Implicit caching runs by default on prompts of at least 4,096 tokens, and cached input costs $0.075 per million tokens. The practical consequence is prompt design. Put the stable material, such as system instructions, tool definitions and reference documents, at the start of the prompt and the variable user content at the end, so repeated requests share a long identical prefix and qualify for the cached rate.

Finally, remember the context window is 1,048,576 tokens, which is large enough to tempt you to paste everything in. Long prompts raise your input bill on every uncached call, so measure before you assume bigger is better. Run a small evaluation on your own data at low, medium and high, and keep xhigh for tasks where an error costs more than the extra reasoning.

Try it today

  1. Set thinking_level in generation_config on every call: low for extraction and classification, medium for drafting, high or xhigh only for hard reasoning.
  2. Replace any thinking level of minimal with low, because Hydrogen returns an error for minimal.
  3. Remove temperature, top_p and top_k from your shared request builder.
  4. Delete explicit cache creation code and restructure prompts so stable content comes first and reaches the 4,096 token threshold for implicit caching.
  5. Track thinking tokens separately in your cost dashboards, because they bill at the output rate of $7.50 per million tokens.
Gemini 4 HydrogenGoogleThinking levelCaching