Gemini 4 Hydrogen: Cheaper Than Argon, and the New Antigravity Orchestrator

Gemini 4 Hydrogen is Google's new "everyday" model in the Gemini 4 family. It launched November 12, 2026, costs $1.50 per million input tokens and $7.50 per million output tokens, runs tasks 2.5x faster than Gemini 3.8 Flash, and slightly edges out Gemini 4 Argon on the Artificial Analysis Intelligence Index (51 vs. 50). It also ships with a new Orchestrator mode in Google Antigravity that hands tasks between Hydrogen and the bigger Gemini 4 Carbon model mid-run.

That's the headline. Now let's get into what's actually true versus what got flattened in the retweet cycle, because a few of the claims floating around (including some I'd have believed at a glance) need a correction.

What is Gemini 4 Hydrogen, exactly?

Hydrogen is the third model in the Gemini 4 line, after Carbon (the heavyweight reasoner) and Argon (the mid-tier coding model). Google's own framing, straight from the Antigravity blog, is that Carbon handles your hardest problems while Hydrogen is "built for rapid execution" — that's a direct quote from Koray Kavukcuoglu, who heads Google DeepMind.

In practice, Hydrogen is a fast, cheap, agentic model meant for the stuff you do all day: writing code, running tests, calling tools, drafting docs, chewing through Gmail and Drive context. It's not trying to out-think Carbon. It's trying to not make you wait.

A few specs worth knowing:

• Context window: 1 million tokens, same as the rest of the Gemini 4 family
• Thinking levels: low, medium, high (default), and a new xhigh tier
• Availability: Gemini API, Google AI Studio, the Gemini app, Google Search, Sheets, Gemini for Business, and paid tiers of Google Antigravity
• Free tier access: yes, via auto model selection when Hydrogen is the right fit for the prompt

That last point matters more than it sounds. Argon never made it to free users in any serious way. Hydrogen did, on day one.

How much does Gemini 4 Hydrogen cost vs. Argon?

Here's the pricing table, and it's where "cheaper than Argon" actually checks out:

So right now Hydrogen is about 25% cheaper than Argon's intro pricing. Once Argon's promo window closes at the end of this month, Hydrogen becomes less than half the price. That gap is the real story, not the benchmark score.

Worth flagging: several HN commenters pointed out that Google has already announced a 2x price bump across its model lineup starting January 1, 2027. So "cheap" here has an expiration date too. Treat today's pricing as a launch promo, not a permanent floor, because that's exactly what happened to Argon.

Is Hydrogen really 2.5x faster than Argon?

No, and this is the correction worth making loudly. The 2.5x figure, which Google DeepMind, Sundar Pichai, and the official Antigravity blog all repeated, is Hydrogen's speed compared to Gemini 3.8 Flash, not Argon. I checked the DeepMind announcement and the Antigravity post directly: both say "2.5 times faster than 3.8 Flash at default settings." Argon doesn't appear in that sentence at all.

That's not a knock on Hydrogen, it's genuinely fast and the Terminal-Bench numbers back it up (57% vs. Flash's 18.2%, more than tripling it). But if you came away thinking Hydrogen smokes Argon on raw speed by 2.5x, that's a headline doing more work than the source material supports. Google never actually published a direct Hydrogen-vs-Argon speed number, just an intelligence score (51 vs. 50, and that's close enough to call a tie once you factor in benchmark noise).

What is Orchestrator in Antigravity?

Orchestrator is the more interesting release here, honestly. It's a routing mode inside Antigravity that stops treating "which model do I use" as a one-time decision. Instead of picking Carbon or Hydrogen at the start of a task and sticking with it, Orchestrator switches between them mid-task based on what the step actually needs.

Think planning and architecture decisions go to Carbon, repetitive execution (writing the function, running the test, fixing the lint error) gets handed to Hydrogen, and the switch happens automatically, in both directions, as the task evolves.

Google's own numbers on Terminal-Bench 4.1:

• Near-Carbon task success rate
• 24% less quota consumed, on average
• 19% faster completion time

There are two flavors: Orchestrator (Carbon) and Orchestrator (Argon), both available to Ultra subscribers, while Pro subscribers get Orchestrator (Argon) only. Varun Mohan (who runs the Antigravity/Windsurf side of things) put it plainly on launch day: Hydrogen is "faster, more efficient, and better than Gemini 3.8 Flash on most evals," and Orchestrator lets Antigravity split work between it and the bigger model automatically.

I think this is the right instinct. The expensive model becomes the planner, the cheap model becomes the worker, and you only pay Carbon prices for the fraction of the task that actually needs Carbon-level judgment. That's basically how a senior engineer delegates to a junior one. The catch, and someone online made this point sharper than I could: more worker agents for the same budget also means more agents carrying the same API keys and repo access. Scaling out cheap agents is a cost win and a blast-radius problem at the same time.

What are the benchmark numbers, actually?

Google also skipped something on the way to shipping Hydrogen: the Fairwind Program, its internal cybersecurity red-teaming process that delayed Argon's release. That's a notable choice to make public. It probably means Hydrogen is a smaller, lower-risk model that didn't need the same scrutiny, but it's also the kind of detail that's worth watching if you're putting Hydrogen in front of untrusted input or giving it tool access.

What are developers actually saying about it?

The reaction split roughly three ways once people started using it instead of reading the press release.

First, the naming scheme is getting mocked. Hydrogen, Argon, Carbon, periodic table elements with no obvious size or capability ordering. One commenter's line stuck with me: stop a random person on the street, ask them to rank hydrogen, argon, and carbon by mass, and most couldn't do it. Fair. OpenAI's numbers and Anthropic's model tiers at least imply a hierarchy you can guess at.

Second, people are skeptical of the "cheap and better" framing holding up under real workloads. One developer said they still "use Opus to review Argon because Argon goes and builds things and also creates a bunch of bugs along with it." That's a complaint about Argon, not Hydrogen specifically, but it's a reasonable question to ask about any model Google ships fast: does the benchmark score survive contact with your actual codebase?

Third, and this is a smaller but sharper complaint, some Antigravity CLI users reported account suspensions with opaque appeal processes, with one person noting the risk of losing a Gmail account entirely over an Antigravity flag. If that's accurate even some of the time, it's a real reason to be cautious about pointing Hydrogen at a Google account you can't afford to lose.

Should you switch to Gemini 4 Hydrogen?

If you're already paying for Argon and mostly doing repetitive coding or agentic tool-calling work, yes, switch, at least until November 30. You're getting a model that scores slightly higher on intelligence benchmarks, costs 25% less right now, and will cost less than half as much once Argon's price jump lands December 1.

If you need Carbon-level reasoning for architecture decisions or gnarly bugs, don't expect Hydrogen to replace it. Use Orchestrator instead and let Google's routing do the handoff for you. That's the actual pitch here, not "Hydrogen beats Argon," but "you probably don't need Argon-or-Carbon as a binary choice anymore."

I'd hold off on calling this a clean win until the January 2027 price hike actually lands. A lot of "cheap new model" launches look great in week one and less great once the intro pricing expires. We saw that exact pattern with Argon five months ago.

Watch the announcement

If you're tracking the rest of the Gemini 4 rollout, our breakdowns of Gemini 4 Carbon and Gemini 4 Argon vs. Claude Opus cover the two models Hydrogen now sits between. And if you want the baseline Hydrogen is benchmarked against, we've also got a full rundown of Gemini 3.8 Flash.