Gemini 4 Hydrogen Scores 51 on Artificial Analysis Index, Edging Past Argon’s 50

TL;DR

• Google released Gemini 4 Hydrogen on November 12, a fast, low-cost tier that sits below Carbon and Argon in the Gemini 4 family.
• It scores 51 on the Artificial Analysis Intelligence Index at default high thinking, one point above Argon’s 50, and 54 at xhigh.
• Google reports it runs 2.5x faster than Gemini 3.8 Flash, lifts Terminal-Bench 4.1 from 18.2% to 57%, and cuts fabricated claims more than fivefold.
• Pricing is $1.50 per million input tokens and $7.50 per million output tokens across a 1 million token window. The figures are Google’s own.

Hydrogen Gets the Grunt Work

Google released Gemini 4 Hydrogen on November 12, and it isn’t pitched as the smartest model in the family. It’s the worker. The target is the agent loop: writing code, running tests, verifying changes, then doing it again.

According to the source file on GitHub, Hydrogen completes tasks 2.5 times faster than Gemini 3.8 Flash at its default high thinking level. Terminal-Bench 4.1 performance rises from 18.2% to 57% against that same Flash model. Fabricated claims fall by more than fivefold.

Pricing is flat across the whole 1 million token context window: $1.50 per million input tokens and $7.50 per million output tokens. On the Artificial Analysis Intelligence Index, Hydrogen scores 51 at default high thinking and 54 at xhigh. Higher-tier Argon scores 50.

The source describes the division of labor in two lines. Carbon coordinates, and Hydrogen runs as worker and verifier agents in isolated workspaces. The other tiers “keep the long-horizon reasoning work, Hydrogen takes the loop that writes code, runs tests and verifies changes.”

A Worker Model That Outscores the Tier Above It

One point. That’s the gap between Hydrogen’s 51 and Argon’s 50, and nobody should pretend it’s a landslide. An index is a single number, and a single number hides what each model is actually good at. But the direction matters. A lower tier landing above a higher one on a third-party index is awkward for any pricing ladder.

Think of a restaurant kitchen where the line cook posts a better tasting-exam score than the sous chef, yet still gets handed the chopping while the head chef calls the orders. The score doesn’t change the job. It changes the argument about who deserves the better station.

The real design choice sits in the routing. Google says the orchestrator in Google Antigravity routes between Carbon and Hydrogen dynamically in a single session. So the pitch isn’t “replace your big model.” It’s “stop paying big-model rates for the loop.” Agent loops chew through tokens, and output tokens here cost $7.50 per million, five times the input rate. If most of a coding session is write, test, verify, repeat, the cheap tier carries most of the bill. Does that hold in practice? Nobody outside Google has said yet.

I’d be careful with the Terminal-Bench number. Going from 18.2% to 57% is a jump over Gemini 3.8 Flash, not a comparison with Argon or Carbon. It tells you Hydrogen is a big step up from its predecessor. It doesn’t tell you it matches the tiers above on terminal work. And the source supplies no Argon figure at xhigh, so Hydrogen’s 54 has nothing to sit beside.

My read: the headline is less about a 51 and more about Google admitting the top model shouldn’t be doing everything. That’s a sensible admission. It’s also a convenient one when the cheaper model is the one you want people to run all day.

Where Hydrogen Sits in the Gemini 4 Family

Hydrogen joins Gemini 4 Carbon and Gemini 4 Argon, both released earlier. It sits below them in the lineup, which makes the index result the odd part of the story. The ranking by tier and the ranking by score don’t line up.

The comparison Google leans on most is Gemini 3.8 Flash. Against it, Hydrogen runs 2.5 times faster at default high thinking, and the reported Terminal-Bench 4.1 and fabrication numbers all use that baseline. That’s a fair way to show generational progress. It’s a less useful way to show how Hydrogen compares with the models a team might actually weigh it against today.

Context matters on what’s missing, too. The material doesn’t include Carbon or Argon pricing, so the cost gap between tiers can’t be calculated here. It doesn’t include independent reproduction of any benchmark. And the source is a news file in a GitHub repository, with every figure traced to Google.

The 1 million token window is worth a sentence on its own. Flat pricing across that whole window means the per-token rate doesn’t change as an agent session stretches out. For long-running loops, that’s the kind of detail that decides whether a budget survives.

Three Things to Check Before Rewiring Your Agents

First, look for independent runs of the Artificial Analysis score and the Terminal-Bench 4.1 result. Google’s numbers are a starting point. A 57% that holds up outside Google’s own harness would be a real story; one that doesn’t would be a footnote.

Second, watch how Antigravity’s routing behaves in real sessions. Dynamic switching between Carbon and Hydrogen only saves money if the orchestrator hands the right work to the right model. Developers will notice quickly if verification tasks get misrouted.

Third, track whether Argon’s and Carbon’s pricing or positioning shifts in response. When a lower tier edges past a higher one on an index, someone has to explain what the higher tier is for. Google’s answer so far is long-horizon reasoning. Whether buyers accept that is the open question.