Gemini 4 Hydrogen Reaches Firebase AI Logic and Rejects Old Settings

Remove thinkingBudget and explicit caching before switching a Firebase AI Logic app to Gemini 4 Hydrogen, or its requests fail with a 400 error.

Firebase AI Logic apps can now call Google’s new gemini-4-hydrogen model, but an app that only swaps the model name may start failing. Requests that set thinkingBudget now return a 400 error, and the default thinking level rises from medium to high. Firebase listed the changes in its Nov. 12 release notes (Firebase release notes, Nov. 12, 2026), the same day Google made the model generally available in the Gemini API.

Key Takeaways #

• thinkingBudget is not supported on Hydrogen. Firebase said there is “no backwards-compatibility” and the request fails with a 400 error.
• The default thinking level is HIGH , up fromMEDIUM ongemini-3.8-flash . Firebase said that may raise latency and token use unless the app setsthinkingLevel toMEDIUM orLOW .
• Explicit context caching is gone. Only implicit caching runs.
• On the Agent Platform Gemini API (formerly Vertex AI), Hydrogen is a short-term availability model.

What Changes #

Firebase said Hydrogen works from client apps through both the Gemini Developer API, including its free tier, and the Agent Platform Gemini API. Its notes for teams moving off gemini-3.8-flash or other Gemini 3.x models list these differences:

• Thinking: AI Logic supports LOW ,MEDIUM andHIGH . The newXHIGH level that Google offers in the Gemini API is not available through AI Logic yet, andMINIMAL still fails with a 400 error, as it did on 3.8 Flash.
• Caching: implicit caching is on by default for prompts of at least 4,096 input tokens on the Gemini Developer API and 6,144 on the Agent Platform API. Explicit caching is not supported.
• Limits: the input limit of 1,048,576 tokens and the 65,536-token output limit match 3.8 Flash. AI Logic does not yet support Hydrogen’s “long decode” mode for longer outputs.
• Parameters: temperature ,topP andtopK are ignored.frequencyPenalty ,presencePenalty andcandidateCount fail with a 400 error.
• Files: Hydrogen and other Gemini 4.x models accept Microsoft Office and OpenDocument files in addition to PDFs and plain text.

Google’s Nov. 12 Gemini API changelog prices Hydrogen at $1.50 per 1 million input tokens and $7.50 per 1 million output tokens on the paid tier. Google’s migration guide says thinking tokens are billed as output. It also tells teams to “test medium and high on your evaluation suite to balance speed and cost.”

Netlify added the model to its AI Gateway and Agent Runners the same day, its Nov. 12 changelog said, so Netlify Functions can call it without managing API keys.

The Take #

The 400 errors show up in the first test run. The thinking default is harder to spot: an app that never set a thinking level was getting medium on 3.8 Flash and gets high on Hydrogen, with no code change and no error. Because Google bills thinking tokens as output, the change shows up as slower replies and a larger bill, not as a failed request. Set thinkingLevel explicitly before switching, then compare the two levels on production prompts.

The short-term availability label matters most for mobile teams. A model name hard-coded in an app on the Agent Platform API can be retired faster than users update the app. Firebase’s own advice is to serve the name from Remote Config or server prompt templates. This is the third Gemini model change Firerun has covered in two weeks, after the Nov. 17 speech preview shutdown and the Dec. 1 Gemini 2.5 reroute. Teams still shipping the model name in client code should move it out now.