Gemini 4 Hydrogen vs Gemini 4: There Is No Gemini 4, and That Is Now a Practical Problem

G​oogle's model reference lists exactly three models under the G​emini 4 heading — Gemini 4 Carbon, Gemini 4 Argon and, since 12 November 2026, G​emini 4 Hydrogen — and not one of them answers to the identifier gemini-4. G​emini 4 Hydrogen is the one that makes the ambiguity expensive. It is the cheapest of the three, the only one with a free tier, the only one of the generation G​oogle ships to Search users worldwide, and the model the Antigravity Orchestrator reaches for when it wants a task finished rather than reasoned about. So "G​emini 4" now most often means Hydrogen, and the teams that mean Carbon need to start saying so in writing.

Three models, three rate cards, one name

• G​emini 4 Hydrogen — generally available 12 November 2026, $1.50 input / $7.50 output per million tokens, flat across the full 1M window, no dated increase announced, free tier available.

• Gemini 4 Carbon — announced 28 October 2026 and generally available 10 November 2026, $2.00 / $10.00 on an introductory rate that G​oogle's rate card ends on 30 November 2026, when both figures double to $4.00 / $20.00. No free tier.

• Gemini 4 Argon — announced 30 September 2026, generally available 13 October 2026, $2.00 / $10.00 on the same introductory rate and the same 1 December step, no free tier.

All three run a unified 1,048,576-token window with long decode up to the same 1M, and all three return text only. That uniformity is exactly why the name is confusing: on a spec sheet they look like one model sold at three prices. On a benchmark they do not.

What the independent scores say the tiers actually are

Artificial Analysis runs its own harness, and its capture puts the three far enough apart that "pick the cheaper one" stops being a strategy.

• Gemini 4 Carbon at high — Intelligence Index 60, 76 output tokens per second, 67.68 seconds to first answer token, $4.00 / $20.00 in its current listing.

• G​emini 4 Hydrogen at xhigh — Index 54, 198 output tokens per second, 41.11 seconds to first answer token, $1.50 / $7.50.

• Gemini 4 Argon at high — Index 50, 64 output tokens per second, 39.90 seconds to first answer token, $4.00 / $20.00 in the same listing.

Read those together and the family stops making sense as a ladder. Hydrogen outscores Argon by four index points, runs about three times faster, and costs less than half as much per token. Argon is not a step up from Hydrogen in any dimension a buyer cares about — it was simply first, and it now occupies the position of a general-purpose model that the newer, cheaper one has walked past.

Carbon is the only genuine ceiling: six index points and roughly 2.7x the price. The gap is real, but it is not the gulf the price suggests, and Artificial Analysis's own per-task figure tells the story — $3.16 of inference to complete one Intelligence Index task for Hydrogen against $7.76 for Carbon. Carbon costs more than twice as much to finish the same task.

The one vendor claim with operational content

G​oogle's launch post is a wall of unreproduced numbers, and the one worth carrying forward is not a benchmark score. Run through Antigravity's Orchestrator on Terminal-Bench 4.0, G​oogle reports Hydrogen reaching near-Carbon task success rates while consuming 24% less quota and finishing 19% faster. That is not a claim that Hydrogen is as smart as Carbon — it is a claim that Hydrogen is nearly as likely to finish the job. For agent work those are different properties, and the second one is usually the one on the invoice.

The only thing that behaves like a single "G​emini 4"

There is one place where the generation acts like one product, and it is not an endpoint. Antigravity's Orchestrator swaps between Carbon and Hydrogen mid-session: Carbon plans and resolves roadblocks, Hydrogen runs the worker and verifier turns. G​oogle's docs now describe that pattern as something you can reproduce yourself — route planning turns to gemini-4-carbon and execution turns to gemini-4-hydrogen inside one multi-turn session using previous_interaction_id, or by preserving thought steps in stateless mode.

That is the closest thing to a real answer to "what is G​emini 4". Not a model — a routing policy. A strong planner handing pieces to a fast executor. If your mental model of the generation is one system you call once, the architecture G​oogle shipped disagrees with you.

Where Hydrogen is the wrong "G​emini 4"

Three cases, and none of them is exotic. The first is anything where a wrong answer is expensive: Carbon's six index points are the difference between a model that needs a verifier and one you can let plan. If you are already paying a second model to check the first, Hydrogen's price advantage is being spent on the checker.

The second is interactive latency, and here the family is uniformly bad rather than neatly ordered. Hydrogen starts answering after 41.11 seconds, Argon after 39.90, Carbon after 67.68 — and 3.8 Flash after 17.10. Anyone streaming tokens to someone watching a spinner is on the wrong generation, whichever G​emini 4 they picked.

The third is recency. G​oogle publishes no knowledge cutoff for Hydrogen on its model card or in the migration guide. If your workload turns on facts from the last few months, that omission is a reason to test before committing rather than a footnote to skip.

Keeping two of them callable

The pattern that survives contact with a real budget is the one G​oogle built into Orchestrator, minus the vendor lock on the policy: Hydrogen for high-volume execution, Carbon for planning and escalation. That is two model families, two sets of quotas and two failure modes for one feature, which is the actual cost of the architecture G​oogle is recommending.

One key collapses most of it. Gemini 4 Carbon and Gemini 4 Argon are both in the OrcaRouter catalogue behind a single endpoint, at provider list price with 0% markup — so when the introductory rate doubles on 1 December 2026, the new figure is live here the same day rather than at the next contract renewal, and switching a route from one to the other is a string rather than a migration. G​emini 4 Hydrogen is not in our catalogue; for that one, G​oogle's own API and AI Studio are the route.

Composition is where it earns its keep with a three-tier family. The routing DSL lets you express a planner-and-executor pair as one call rather than inheriting someone else's policy, and model fusion — a panel answering together — is the honest alternative to guessing which tier a bare "G​emini 4" meant. Neither is a substitute for deciding. Both make the decision reversible, which matters when two of the three models are three weeks from a price change.

What to actually call it

If someone asks for G​emini 4, ask which one, and expect the answer to be Hydrogen now. The generation has a ceiling (Carbon), a workhorse (Hydrogen) and a first draft (Argon), and the price ordering no longer matches the capability ordering — Argon costs almost what Carbon does and scores four points below Hydrogen. That is a model priced for a market that stopped existing on 12 November 2026.

Two datable things will settle the naming faster than any explainer. Whether G​oogle publishes a knowledge cutoff for Hydrogen, and whether the 1 December step on Carbon and Argon holds — a $4.00 / $20.00 Carbon against a $1.50 / $7.50 Hydrogen makes the tier choice trivial for anyone who can tolerate the score gap, and G​oogle's own introductory-rate language says the change is subject to revision. Until either lands, the correct reading of "G​emini 4" is a family, and the correct habit is to name the model you mean.