Skip to main content

Gemini 4 Hydrogen: what to check before replacing 3.8 Flash

Gemini 4 Hydrogen is Google's new fast Gemini 4 model, open on every Gemini API tier. What changed from Gemini 3.8 Flash, the output and refusal limits in the model card, and the checks to run before you swap it into an agent.

Share -
Developer in a checked shirt, seen from behind, working at a laptop showing code beside a glowing blue neural network brain, in a dark room lit by a cyan lamp

Gemini 4 Hydrogen is Google's fast, lower-cost Gemini 4 model for everyday coding and agent work, adapted from Gemini 4 Carbon and open on every Gemini API tier. It is the natural replacement for Gemini 3.8 Flash, but test refusals, parallel tool calls and output length on your own workload before you switch.

What was released

Gemini 4 Hydrogen was announced by Google on 12 November 2026 and is rolling out globally the same day. Google positions it as the execution model of the Gemini 4 family: Carbon and Argon handle deep, long-horizon reasoning, and Hydrogen handles fast, repeated work such as writing code, running tests and checking changes. Access by channel:

  • Gemini API and Google AI Studio: all tiers, according to the announcement. The AI Studio link in the announcement uses the identifier gemini-4-hydrogen, and Google's developer post points to a what's new guide in AI Studio for setup and thinking level controls.
  • Android Studio: available for building with the Gemini API.
  • Gemini for business: through the API, the model dropdown, Google Antigravity and Gemini in Google Sheets for eligible business subscriptions.
  • Consumers: the Gemini app model dropdown and Google Search for all users, Gemini in Sheets from Google AI Pro and Ultra, and Antigravity for paid Pro and Ultra subscribers.

Google also introduced Orchestrator in Antigravity, which switches between Carbon and Hydrogen within one session. That is an Antigravity feature, not an API capability.

What changed from Gemini 3.8 Flash

Google describes Hydrogen as a standalone upgrade for teams moving off Gemini 3.8 Flash. The changes that matter for an engineering team:

  • Four thinking levels: low, medium, high (the default) and a new xhigh tier, so you can trade quality against latency and cost per call.
  • Flat pricing across the context window: one input rate and one output rate across the whole 1M-token window, with no higher tier for long prompts. Google lists the rates in the announcement.
  • Parallel tool calls: Google says the model batches file-read and search calls in parallel instead of inspecting files one by one, and recovers when a tool call fails or a search returns nothing.
  • Fewer fabricated claims and less overreach: Google reports fewer made-up claims than 3.8 Flash and fewer attempts to bypass permission boundaries when a task is blocked.

The Gemini 4 Hydrogen model card adds the limits:

  • Text, image, audio and video input within a combined input and output context of up to 1M tokens.
  • Up to 64K output tokens by default for a synchronous call; longer outputs, up to the full window, need background long decode.
  • Knowledge cutoff of March 2026.
  • Unjustified refusals are higher than on 3.8 Flash, particularly for multimodal prompts.
  • Occasional slowness or timeouts, and higher token use at higher thinking levels.
  • The full safety and security report is due in the coming days.

Benchmark figures in the announcement are Google's own. Treat them as a reason to test, not as a result.

What it means for teams building AI agents

Run your 3.8 Flash eval set first. Hydrogen is a new model, not a patch, so behaviour on your prompts will move. Run the same test set you use for 3.8 Flash and compare pass rate, tokens per task and latency at each thinking level. Keep the lowest level that passes. Our notes on what AI agent evals catch cover how to build that set.

Watch refusals on image and document flows. The model card's one stated regression is more unjustified refusals, especially on multimodal prompts. If your agent reads invoices, screenshots or scanned forms, add those cases to the eval set and track refusal rate on its own, not only overall accuracy.

Make your tools safe to call in parallel. A model that batches tool calls will hit your file, search and database tools at the same time. Check that handlers are idempotent, that rate limits on downstream APIs can absorb a burst, and that a failure in one call returns a clear error the model can act on. The patterns in handling a failed AI agent tool call apply directly.

Check output length. Agents that generate long reports or large code files in a single synchronous call may hit the 64K default and need the background long decode path instead. Find those flows before switching, not after.

Use it as the worker, not the planner, when the task is hard. Google's own pattern is Carbon or Argon as coordinator and Hydrogen as worker and verifier. For multi-agent systems, that split keeps the expensive model on planning and review while Hydrogen does the volume. Budget it using the approach in controlling AI agent running costs.

When not to switch yet. Wait if multimodal refusals already cause support tickets on 3.8 Flash, if your compliance team needs the full safety report before approving a model, or if you rely on Vertex AI, which the announcement does not name as a channel. Plan the move either way: our model deprecation notes explain why keeping the model ID in configuration makes the change cheap.

Frequently asked questions

What is the model ID for Gemini 4 Hydrogen?

The announcement's Google AI Studio link uses gemini-4-hydrogen, which points to that identifier for the Gemini API. Google's what's new guide in AI Studio gives setup instructions and thinking level controls. Confirm the exact ID in the AI Studio model list for your project before you hard-code it, and keep it in configuration so you can move between Hydrogen, 3.8 Flash and Carbon without a code change.

Is Gemini 4 Hydrogen available on the free tier of the Gemini API?

Google says Hydrogen is available for all tiers of Google AI Studio and the Gemini API, which includes the free tier. Free tier rate limits and data use terms still apply, so move production traffic to a paid project before relying on it for customer-facing work, and check that your quota covers the burst of parallel tool calls the model makes.

What are the thinking levels in Gemini 4 Hydrogen?

Hydrogen offers four thinking levels: low, medium, high and a new xhigh tier. High is the default. Higher levels spend more tokens on reasoning, which raises cost and latency, and the model card warns that token use grows at higher settings. Test each level on your own tasks and pick the lowest one that passes, rather than leaving every call at the default.

How long can a Gemini 4 Hydrogen response be?

According to the model card, a synchronous call returns up to 64K output tokens by default. Longer outputs, up to the full 1M-token context window shared between input and output, need background long decode. Agents that write long documents or large files in one call should be checked against the default limit before switching, and moved to the background path where needed.

Share this post

Contact us to build the right product

Talk to our engineers about your application, the systems it connects to, and what you want to build next.

Get in touch
Two people discussing work with a laptop