Gemini 4 Hydrogen: Google's cheaper agent model scores 54 on an independent index, and talks a lot

Gemini 4 Hydrogen is Google's new workhorse. Google DeepMind released it on November 12, 2026 as the fast, lower-priced tier under Gemini 4 Carbon and Gemini 4 Argon, and the first independent numbers suggest it is capable and cheap per token, but not necessarily cheap per task. Here is what Google claims, what Artificial Analysis measured, and what the model card admits.

What Google shipped

Koray Kavukcuoglu, Google DeepMind's chief AI architect, announced Hydrogen in a post on Google's blog dated November 12. Google positions it for coding, debugging and agent workflows, and says Carbon and Argon remain the models for deep, long-horizon reasoning. It has four thinking levels: low, medium, high (the default) and a new xhigh tier. The context window is 1 million tokens. The model card lists a knowledge cutoff of March 2026 and says Hydrogen was adapted from Gemini 4 Carbon.

Access is broad on day one. Developers get it through the Gemini API, Google AI Studio and Android Studio. Consumers get it in the Gemini app's model picker and in Google Search, and paying Google AI Pro and Ultra subscribers get it in Google Antigravity, Google's coding environment. Google also introduced an Orchestrator in Antigravity that hands planning to Carbon and routine work to Hydrogen. On Terminal-Bench 4.0, Google says the pairing reached near-Carbon task success with about 24% less quota and about 19% faster completion.

The price

Hydrogen costs $1.50 per million input tokens and $7.50 per million output tokens, with no surcharge for long context. Artificial Analysis lists the median for comparable models at $2.00 and $10.00, so Hydrogen sits about a quarter below the middle of the market on both. For teams already running Gemini 4 Carbon, the cheaper tier is the obvious thing to test next.

What the benchmarks say

Google's own table is favorable. It reports a score of 51 on the Artificial Analysis Intelligence Index at the default setting, one point above Argon's 50 in Google's comparison, and 54 at xhigh. It reports 57% on Terminal-Bench 4.1 against 18.2% for the previous Gemini 3.8 Flash, 58.8% on HLE-Diamond, and first place on the Vals tax agent benchmark at 52.3%.

Artificial Analysis, an independent evaluator, confirms the headline number. It scores the xhigh setting at 54, ranked 16th of the 78 models on its list, against a median of 41. That is a real result. The Terminal-Bench and Vals figures are Google's own reporting and we have not seen independent replications of them as of November 13.

The same Artificial Analysis page is where the picture gets less tidy. Hydrogen generated 304 million output tokens to finish the index, against a median of 72.8 million, which ranks it 76th of 77 for verbosity. Time to first answer token was 40.55 seconds against a median of 13.61, though output speed once it starts is quick at 188 tokens per second. The cost to run one index task came to $3.16, and a task took 13.7 minutes against a 13.0 minute median. In plain terms: a model priced below the median ends up around the median on cost per finished task, because it thinks at length.

Google says partners saw 12% to 23% fewer tokens than Gemini 3.8 Flash on enterprise workflows. That claim is about a predecessor on partner data, not about the market on a public index, and the two do not contradict each other.

What the model card admits

The Frontier Safety Framework section is the part most coverage will skip. Google says Hydrogen showed no material capability increase over Argon, so it did not run the full set of critical capability evaluations. Targeted testing found it does not reach the critical capability levels for CBRN (chemical, biological, radiological and nuclear) or cyber, but it does reach the alert thresholds for both, and mitigations stay in place. It also did not reach the levels for harmful manipulation or for accelerating machine learning research.

For loss of control, Google did not test Hydrogen against its stealth and situational awareness level because capability had not changed from Argon. As a conservative step, it applied the mitigations as if that level had been reached. The full Safety and Security Report is listed as coming soon, which means the strongest safety numbers in the launch post, such as a 12x cut in misaligned actions and 96.7% honesty on an internal evaluation, cannot be checked yet. For how these frameworks differ, see our comparison of the major labs' safety frameworks.

The card also lists limits plainly: hallucinations, jailbreak resistance still being improved, occasional slowness or timeouts, and higher thinking levels that use more tokens.

Who should care

Teams running agents at volume are the audience. The pattern Google is selling, a big model that plans and a cheaper one that executes and verifies, is where agent spending is heading, and a one million token window at a flat rate helps with long repository work. The risk is the token appetite. A high-effort run on a verbose model can erase a per-token discount, and a 40 second wait for the first token is a poor fit for anything a user is watching.

Buyers should also note that this is a vendor-first launch. Most of the benchmark wins are Google's own, the independent index covers one provider route, and the safety report has not landed.

Our take

Hydrogen is a credible, well-priced model, and the independent 54 is the number that matters. Do not read the $1.50 as the cost of your workload. Run your own tasks at the default and xhigh settings, log tokens and finished-task cost, and compare against Argon before switching. We will update this piece when the Safety and Security Report is published.