Skip to content

Gemini 4's Three Names Are a Division of Labour, Not a Leaderboard

Carbon, Argon and Hydrogen read like a quality ladder. A Japanese AI director who uses them daily argues they are roles: one plans, one executes. His numbers are worth reading with care, and one of them matters more than the rest for anyone generating images or video at volume.

AI-assisted draft. Reviewed and edited by the Phosphene team before publication.

On this page
A heavy brass drafting compass and a rolled plan on the left of a workshop bench, a row of small machined gears on the right, one turning under a hard shaft of light.
Editorial illustration, not a product screenshot.

Google now ships three Gemini 4 tiers with element names, and the naming is doing damage. Carbon, Argon, Hydrogen in that order looks like a ladder: heavy at the top, cheap at the bottom, pick the one you can afford. ZEN, an AI director in Japan who writes up his tooling on note.com, published a hands-on comparison on 14 November that argues the opposite. The three are not ranks. They are jobs.

His framing, translated: Carbon and Argon are the supervisors who think through a plan, Hydrogen is the worker who executes a defined task. Using a thinking model for routine work does not buy quality, it buys latency. The claim that Hydrogen is not simply the weak one holds up beyond his post: Google puts it at 51 on the Artificial Analysis Intelligence Index at its default thinking level, against 50 for Argon, and 54 at its highest setting.

Treat the rest the way you should treat all of it. The speed multiplier is Google's, measured against the Flash model Hydrogen replaces rather than against Carbon or Argon. The overreach and invention figures are ZEN's own observations from daily use, with no published method behind them. He also sells coaching and tooling, which is a reason to read the numbers as observations rather than measurements. The structure of his argument holds up independent of the exact figures, and that structure is the useful part.

The three numbers worth remembering

The changes he reports are not the usual ones.

Google's headline speed claim is 2.5 times faster task completion than Gemini 3.8 Flash, with Hydrogen running at its default high thinking level, and ZEN reports the same order of change in daily use. His own figures point the same way: overreach, his word for doing more than you asked, drops by roughly 80 percent, and unsupported invention falls to under a fifth.

The output ceiling needs more care than it usually gets. Google documents a 65,536-token default output limit for a single synchronous request on all three tiers, with Carbon and Argon capped at 32,768 for non-streaming direct HTTP calls. Hydrogen's advantage sits in that mode. Long decodes extend any of the three to the shared million-token window, so nobody is saved from truncation by picking a different element.

Of those, the overreach figure is the interesting one for anyone working with generative media. Speed is a convenience. Obedience is a behavioural property, and it is the one that quietly ruins batches.

Why less overreach beats more intelligence

Think about the actual failure of a fast model in a production loop. It almost never fails by being stupid. It fails by being eager. You ask it to shorten one line in a prompt and it rewrites the whole prompt, removes the lens description you spent an hour on, and helpfully adds "cinematic lighting". You ask for ten captions and get ten captions plus an introduction and a closing paragraph you now have to delete.

At a single request that is annoying. Across a batch of forty prompt variants it is the difference between a usable afternoon and a redo. This is the same reason people stop using agents for mechanical edits: the work of reviewing the extra work exceeds the work itself.

If the claim survives your own testing, it changes what you should route to the cheap tier. Not just "simple questions". Specifically the tasks where the instruction is already fully specified and any creativity is a defect: reformatting a shot list into a table, translating a prompt into your house style without changing the content, extracting the fifteen prompts from a planning document, renaming files from a naming convention, turning a call transcript into action items.

Where each tier belongs in an image or video pipeline

The trap is thinking of this as a chat-model question. It is a pipeline question, and the pipeline has three kinds of step.

Planning steps want the heavy model. Deciding the shot order for a fifteen-second sequence, diagnosing why a render came back with the wrong anatomy, working out whether a failure is the prompt, the reference image or the seed. These are the steps where being wrong costs a batch, so latency is cheap.

Execution steps want the fast model. Bulk prompt rewriting, caption generation, transcribing a client's brief into fields, pulling the reusable half of a prompt library into a new template. The work is defined, the volume is high, and you will read the output anyway.

Judgement steps need a different arrangement, and this is where the routing error costs the most. All three tiers take image input, so the question is not whether they can see a frame. It is whether a model reasoning about twenty candidates is doing the same job as a person glancing across a contact sheet. Test it on your own material before you assume either answer: hand the tier a grid of variants, ask which three survive, and count how often you agree. A model that confidently rejects the frame you would have printed is worse than no opinion, because it costs you a review and gives you nothing back.

There is a fourth kind of step that people skip, which is the one that decides everything else: writing down which category each step is in. Ten minutes with a list beats an afternoon of model shopping. It is the same discipline as picking a video model by task rather than by leaderboard, which is what we argued in the breakdown of Wan 3.0's cost tiers.

The split is being built into the tools

The most concrete detail in ZEN's post is not a benchmark. It is that Google's Antigravity developer tool now ships with an official pairing where Carbon plans and Hydrogen executes and runs tests. That is a vendor describing the same division of labour in its own product, which is a stronger signal than any single review.

It also explains why the names confuse people. The tiers were not designed to be chosen once. They were designed to be swapped within a session, and most interfaces still present them as a dropdown with a price next to each option. A dropdown implies a permanent decision. The work implies a handoff.

What to ignore, and what to test

Ignore the leaderboard margin. One or two points on a reasoning index tells you nothing about whether a model will preserve your prompt's grammar while reformatting it, which is the property your batch actually depends on.

Test the thing you actually care about, with your own material. Take five real tasks from last week, run each on both tiers, and count manual repairs instead of scoring impressions. Two numbers come out: how many edits the output needed, and how long you waited. If the fast tier needs the same number of edits, the heavy tier is only buying you confidence. If it needs more, you have found the boundary, and the boundary is worth more than any ranking.

One more caution, because this is where these comparisons go wrong. Two of the claims in the post are about what competing models cannot do: whole-file audio and video ingestion on a million-token budget, and stable quota consumption when you pass a lot of material at once. Those are worth verifying against current documentation before you build a workflow on them, since ingestion limits are exactly the kind of thing that changes between releases.

The practical takeaway is dull and it survives the next model launch. Stop asking which model is best. Ask which step you are standing on, and whether this step needs a decision or an execution. Then pick the tier that is bad at the other one.

Sources

Share this guide
X LinkedIn