Gemini 4 Hydrogen Makes Speed a First-Class Model Choice

Google is framing Gemini 4 Hydrogen as an execution model for everyday coding and agent workflows, rather than its deepest reasoning option. Its model card says Hydrogen is adapted from Gemini 4 Carbon and takes text, images, audio, and video with a combined input-output context window of up to one million tokens. Google’s announcement says its default high thinking level completes tasks 2.5 times faster than Gemini 3.8 Flash, and reports a Terminal-Bench 4.1 score of 57%, compared with 18.2% for Flash. Those are provider-reported comparisons, so they describe Google’s evaluation, not an independent reproduction.

The product choice is notable because Google presents Hydrogen both as a standalone model and as a worker that can be paired with larger models. Its announcement describes an Antigravity Orchestrator that switches between models, and says a Terminal-Bench 4.0 test reached near-Carbon task success while consuming 24% less quota and finishing 19% faster. That suggests a practical tradeoff for developers: reserve a higher-capability model for planning or difficult judgment, then route execution work to a faster model. The numbers come from Google’s own test description, however, and the announcement does not provide enough detail in that passage to independently assess task selection or workload representativeness.

The model card gives a more mixed picture than a single headline score. It reports Hydrogen ahead of Claude Sonnet 5.5 and GPT-6.1 Terra on Agents’ Last Exam v1.1, 71.6% versus 60.4% and 69.8%, and on MKA, 53.4% versus 52.0% and 43.8%. But on Terminal-Bench 4.1, the card lists 57.0% against Sonnet’s 63.9%, and on OSWorld-2.1’s offline subset it records 77.6%, below Sonnet’s 83.9%. These are benchmark-specific results assembled by Google; they do not establish a general ranking, and the card links to a separate methodology page for evaluation details.

Hydrogen also gives users four thinking levels: low, medium, high, and xhigh. Google says these let developers choose a balance among speed, reasoning depth, and cost, and prices API use at $1.50 per million input tokens and $7.50 per million output tokens across the context window. That makes it easier to compare the model in a real workload by latency and spend, rather than treating benchmark scores as the whole decision. For teams testing agent systems, the relevant question is whether the lower-cost setting completes the task reliably enough to avoid retries, extra tool calls, or human correction.

There are meaningful caveats in the model card. Google lists hallucinations, occasional slowness or timeouts, and says higher thinking levels may use more tokens; it gives a March 2026 knowledge cutoff. Its automated safety comparisons show a 5.57 percentage-point increase in unjustified refusals versus Gemini 3.8 Flash, particularly for multimodal prompts, while the full safety report is still pending. The card also says the tests use improved evaluations that are not directly comparable with earlier Gemini model cards. So Hydrogen’s speed and task scores are useful starting points, but developers should measure their own error rates, latency, and refusal behavior before routing consequential work through it.