Gemini 4 Hydrogen raises the baseline for everyday agents

What Google released

Google’s launch post [1] positions Hydrogen as the fast execution member of the Gemini 4 family. Google made it generally available through the Gemini API and Google Cloud, and also placed it in Search, the Gemini app, AI Studio, Sheets and Antigravity.[1][7]

The model takes text, images, audio and video as input and produces text. Its one-million-token allowance is shared by the prompt, reasoning, function calls and final response. A normal foreground segment can produce up to 65,536 tokens, while supported background or continuation flows can extend generation within the shared limit.[7][8]

Hydrogen supports function calling, code execution, computer use, structured output and several grounding tools. It does not support the Gemini Live API, tuning or RAG Engine on the Google Cloud model page.[7]

Why the release matters

Hydrogen gives teams a middle option between a low-cost small model and a frontier model used for every step. Google reports 57% on Terminal-Bench 4.1, compared with 18.2% for Gemini 3.8 Flash, although Claude Sonnet 5.5 scores 63.9% in the same Google model-card table.[1][3]

Google Antigravity’s new Orchestrator can move between Carbon and Hydrogen as a task changes. Google reports near-Carbon task success with 24% less quota and 19% faster completion on Terminal-Bench, but this is a vendor evaluation of Google’s own routing system.[2]

The broad rollout also matters beyond developers. Google made Hydrogen available globally in Search, where Google says it improves query fan-out and complex-question handling. Search users may discover a different range of sources even if they never select the model directly.[1]

Cost depends on the whole task

Google charges $1.50 per million uncached input tokens, $0.075 per million cached input tokens and $7.50 per million output tokens on the standard paid Gemini API. Output billing includes reasoning tokens.[4]

Gemini 3.8 Flash keeps an introductory price of $0.75 input and $3.75 output through December 31, 2026. Google currently charges twice as much per uncached token for Hydrogen. Teams should compare completed-task cost, retries and reviewer time rather than treating model speed as a cost result.[4][5]

Teams may also pay for more reasoning tokens. Hydrogen defaults to high and adds xhigh.[8] Google’s model card warns that higher thinking levels may use more tokens. It separately notes that the model can sometimes be slow or time out.[3]

What independent testing found

Plotly ran Hydrogen through a private 43-question analytics exam and reported that it passed every question for $6.36. Hydrogen caught all five planted hidden patterns and all three data-quality problems. Plotly manually checked the passes against its answer key.[6]

Readers should not treat that result as a universal ranking. Plotly says it ran Hydrogen once, and repeat runs of other models moved by as many as three questions, and the harness emphasizes tool-driven data analysis. GPT-6.1 Sol matched the perfect score for $2.61 and finished with a lower median time per question.[6]

VentureBeat tells buyers to ask whether Hydrogen may replace retries or stronger-model calls, but it is materially more expensive per token than several small-model alternatives. The answer depends on your workload and quality threshold.[5]

A migration test for Gemini 3.8 Flash

Run Hydrogen beside the current model before changing production traffic. Developers who change only the model name can hit several API differences documented by Google Cloud.[8]

• Copy 50 to 200 recent tasks with expected outcomes, tool traces and reviewer decisions. Remove secrets and personal data before testing.
• Run Flash and Hydrogen with the same tools, permissions and stop conditions. Test low, medium and high reasoning separately.
• Measure successful completion, incorrect actions, retries, wall-clock time, input tokens, reasoning and output tokens, cache hits and reviewer time.
• Remove thinking_budget and MINIMAL before migration because Hydrogen returns HTTP 400 for those settings. Custom temperature, top_p and top_k values are ignored.[7][8]
• Check long prompts carefully because input, reasoning, tool calls and output share one 1,048,576-token allowance.[7][8]
• Keep a fallback for timeouts, blocked tools and malformed structured output, then canary the winning configuration on low-risk traffic.
• Re-run the matched set before raising the reasoning level or enabling computer use, new grounding tools or automatic actions.

Limits and open questions

Google has not released the model weights or parameter count. The model card says Hydrogen is adapted from Carbon, but it refers readers to Carbon for architecture and training-data details.[3][6]

Hydrogen’s knowledge cutoff is March 2026, and Google warns that some domains may be less current. Grounding and retrieval still need source checks, especially for current events and high-impact decisions.[3]

The full Hydrogen Safety and Security Report was still marked as forthcoming on the launch-day model card. Google says targeted testing did not put Hydrogen above its critical cyber or biological capability levels, although it reached alert thresholds in both domains.[3]

The model card reports a 5.57 percentage-point increase in unjustified refusals versus Gemini 3.8 Flash in an automated evaluation. Teams working with benign but sensitive material should include refusal quality in their test set.[3]

Frequently asked questions

• Is Hydrogen cheaper than Gemini 3.8 Flash?[4] Not at current standard token prices. Hydrogen costs twice as much per uncached input and output token until Flash’s introductory pricing ends, although it may complete some tasks with fewer retries or tokens.
• Should Hydrogen replace Carbon?[1][6] Use task evidence. Hydrogen can handle substantial coding and analytics work, while Google still positions Carbon for difficult planning and long-horizon reasoning.
• Does the one-million-token window mean one million output tokens?[7][8] Only when the prompt and all other output fit inside the same shared allowance, and long output requires a supported continuation or background flow.
• Is xhigh always better?[3][5] No. Higher reasoning may use more tokens, and VentureBeat reports that independent Terminal-Bench 4.0 testing found no gain from high to xhigh on that test.
• Can I use Hydrogen for voice agents?[7] It accepts audio input, but the Google Cloud model page says the Gemini Live API is not supported.