● Short answer: Hydrogen for agents and tool use. Argon for heavy reasoning. The smaller model wins more benchmarks than the bigger one.
● Index: Hydrogen 51 at high and 54 at xhigh. Argon 50.
● Agentic coding: SWE-sweep 13.5% against 4.3%. Not close.
● Where Argon still wins: Chartography 71.6% to 66.6%, plus GraphWalks and SEC-bench Pro where Hydrogen has no published figure.
Google shipped Gemini 4 Argon as its flagship, then shipped Gemini 4 Hydrogen four weeks later and Hydrogen scores higher on the Artificial Analysis Intelligence Index. That is an unusual thing for a vendor to publish about its own lineup, and it is worth understanding before you pick one.
The published numbers
| Benchmark |
Gemini 4 Hydrogen |
Gemini 4 Argon |
| AA Intelligence Index |
51 high, 54 xhigh |
50 |
| SWE-sweep, agentic coding |
13.5% |
4.3% |
| Agents’ Last Exam v1.1 |
71.6% |
69.0% |
| OSWorld-2.1, offline subset |
77.6% |
74.7% |
| Vending-Bench 2, end balance |
$14,549 |
$13,718 |
| Chartography |
66.6% |
71.6% |
| GraphWalks |
not published |
84.2% |
| SEC-bench Pro |
not published |
83.4% |
| Terminal-Bench 4.0 |
not published |
57.1% |
Read the blanks. Three of Argon’s strongest results have no Hydrogen figure beside them. That is not Hydrogen losing those benchmarks — it is Google choosing not to publish them, which usually means the number was not flattering. Treat the bottom three rows as Argon territory rather than as Hydrogen failures.
Price, and the part that changes your bill
Hydrogen is $1.50 per million input tokens and $7.50 output, with cache hits at $0.075. The detail that matters more than the headline: that rate is flat across the entire 1M-token context rather than stepping up past a threshold. Long-context work does not get repriced mid-request.
Artificial Analysis puts cost per task at $0.29 at low effort, $0.80 at medium, $1.87 at high and $3.16 at xhigh. So the xhigh tier that buys you 54 on the index instead of 51 costs roughly 1.7 times a high-effort run. Three index points is rarely worth that.
Which one, concretely
| If you are… |
Pick |
| Running agents or tool-heavy workflows |
Hydrogen. SWE-sweep is three times Argon. |
| Doing deep multi-step reasoning |
Argon. GraphWalks and SEC-bench Pro are its ground. |
| Reading charts and documents |
Argon, 71.6% to 66.6% on Chartography. |
| Watching latency |
Hydrogen. 2.5x faster than 3.8 Flash at default. |
| On a tight budget |
Hydrogen at low effort, $0.29 per task. |
| Already on Gemini 3.8 Flash |
Hydrogen. It triples Terminal-Bench 4.1 and cuts fabricated claims more than fivefold. |
Listings: Gemini 4 Hydrogen. Background: the launch in full.
Sources