Gemini 4 Hydrogen Lands in Antigravity: What the Update Gate Is Really Gating

The least interesting sentence in the announcement is the one that matters most operationally. G​emini 4 Hydrogen does not appear in Antigravity because you asked for it; it appears because you updated. The model shipped on 12 November 2026, and the Antigravity build that exposes it is the same build that exposes Gemini 4 Carbon and Gemini 4 Argon as planning orchestrators, so an out-of-date client is not missing one model — it is missing the whole three-model arrangement G​oogle now treats as the default way to write code. That is why the rollout note reads like a changelog instruction and why it is worth taking literally.

What is actually new, dated, and checkable

Three things changed inside a single day, and only one of them is the model.

• The model. G​emini 4 Hydrogen entered G​oogle's model list and rate card on 12 November 2026 at $1.50 per million input tokens and $7.50 per million output, on a 1,048,576-token window. The Artificial Analysis page for it carries the same release month and an Intelligence Index of 54. That part is a launch, and it is one day old.

• The client. Antigravity's own post announcing the model is dated 12 November 2026 and describes Hydrogen as "a standalone model for everyday coding and agentic work" that can be selected directly or reached indirectly through an Orchestrator dropdown. No version number appears on that page, so "the latest version" is a claim you have to verify inside the app rather than against a published build string. If the model selector does not list Hydrogen, you are on the old build.

• The default. Antigravity's model documentation now lists Hydrogen as available on every plan tier it documents — Free and AI Plus, Pro, Ultra, Enterprise — while the announcement post describes the rollout in terms of paid Pro and Ultra subscribers. Those two statements do not agree, and the honest reading is that the documentation table is where the rollout is heading and the announcement is where it starts. If you are on a free tier and do not see it, that is an expected staging difference rather than a bug.

What has not changed: there is still no model called gemini-4, and there is still no string with "flash" in it for this model. The identifier is gemini-4-hydrogen.

The (Flash) in the rollout post, and why the API has no such word

The post that surfaced this calls it "G​emini 4 Hydrogen (Flash)". That parenthetical is doing real work and it is not G​oogle's. It is the community's shorthand for the slot Hydrogen occupies — the fast, cheap, high-volume member of the family, the job Gemini 3.8 Flash used to do. G​oogle's own product language never uses it: the model list calls Hydrogen the model that brings "cutting-edge intelligence to everyone", the Antigravity post calls it standalone, and the identifier contains no flash token at all.

This matters for exactly one practical reason. If you go looking for the model by pattern — the way you would look for gemini-3.8-flash — you will not find it, and you will conclude the rollout has not reached you. Search the picker for Hydrogen, and request gemini-4-hydrogen in code. The naming change is the first thing that breaks in a migration, and it breaks silently, because a request to a non-existent Flash identifier fails at the string rather than at the model.

Four thinking levels, and the one that is not the default

The picker shows low, medium, high and a fourth level labelled Extra, which maps to the xhigh reasoning budget. High is the default. On the command line the same setting is reachable as --model gemini-4-hydrogen-xhigh or --effort xhigh.

The gap between those settings is the single largest source of disagreement between the numbers you will read this week. G​oogle's Antigravity post quotes 51 on the Artificial Analysis Intelligence Index for Hydrogen, "ahead of Argon's 50, and up to 54 at its highest setting". The independent page reports 54 because it benchmarks the highest setting. Both numbers are correct and neither describes the model you get if you never touch the picker. A team comparing "Hydrogen at 51" against "Carbon at 60" is comparing a default against a maximum, which is an unfair fight in one direction and an unflattering one in the other.

Orchestrator is the exception: G​oogle documents it as running a fixed configuration, so there is no thinking level to choose when Carbon or Argon is doing the planning. The planning turn is not tunable; the execution turn is.

Three ways to reach the same model, and what each one costs you

Inside Antigravity, Hydrogen is reachable three ways, and they are not equivalent purchases.

• Hydrogen directly. You pick it, you pick a thinking level, you pay for its tokens and its thinking. The whole task runs at index 54 behaviour at best, or 51 at default, with no planning pass.

• Orchestrator (Carbon). Carbon plans, Hydrogen executes. The Antigravity team reports this configuration used on average 24% less quota and completed 19% faster on Terminal-Bench 4.1 while landing success rates near Carbon's own. That is a claim about finishing tasks under a quota, not a claim about intelligence, and it is the only claim in the rollout that changes what an agent does rather than what it scores.

• Orchestrator (Argon). The same arrangement with the older planner. Per G​oogle's plan table, Ultra subscribers get both orchestrator variants and Pro subscribers get only the Argon one, which makes Carbon planning a plan-tier feature rather than a model feature.

The quota sentence deserves to be read twice, because "less quota" is the part that shows up on an invoice. Frontier quota is the scarce resource in these tools, and a configuration that spends a quarter less of it to reach a similar success rate is a better deal than a model with a higher index that burns the budget by step four. That is also the argument against running Carbon on every step, which the post makes directly: it "burns frontier quota" for work Hydrogen can do.

The bug this rollout will produce, and how to spot it in a day

Three failure modes are likely enough to pre-empt.

The first is a stale client. The model simply is not in the selector, which is obvious and self-correcting after an update. The second is subtler: an updated client, a working selector, and a request that still routes to the old default because a saved session or a checked-in config pins the previous model. Antigravity lets you save the model per context, so a workspace that was set up in October keeps October's answer until someone changes it. Check the model name that actually ran, not the one you picked.

The third, and the one that costs money rather than time, is thinking-level drift. A workload tuned on high and then moved to Extra for a hard case will look like a regression in throughput and like an improvement in quality at the same time, because it is both. G​oogle's own two published figures — 51 and 54 — are the same model at two settings, which is a useful reminder that any internal benchmark you run this week is a benchmark of a configuration.

Where this leaves teams who route rather than click

The Antigravity path is convenient and it is also a single vendor's client deciding which model runs. For anything that has to keep working when a picker changes or a plan tier is re-cut, the durable version is an API call whose model string you control, with a fallback behind it.

Worth stating plainly, because the rollout posts blur it: G​emini 4 Hydrogen is not in OrcaRouter's catalogue. It is available through G​oogle's own API and AI Studio and, as of this week, inside Antigravity. What the catalogue does carry at provider list price with 0% markup is Gemini 4 Carbon at $2.00 and $10.00 per million, Gemini 4 Argon on the same card, and Gemini 3.8 Flash at $0.75 and $3.75 — all callable from the same key, with automatic failover between providers when one path degrades. The honest configuration for a team migrating this week is Hydrogen called direct for the execution turn, with Carbon or 3.8 Flash behind the router as the path a request falls back to when the new model throttles, a region misbehaves, or a plan tier turns out not to include what the documentation table promised.

That split also survives the calendar. Carbon and Argon double on 1 December 2026, from $2.00 and $10.00 to $4.00 and $20.00 per million. Hydrogen's published card carries no dated step at all. If that asymmetry holds, the tier structure G​oogle is building is a $4/$20 ceiling and a $1.50/$7.50 workhorse, and the workhorse is the one that just appeared in your editor.

What to check before you trust the rollout note

• Confirm the build, not the blog. Open the model selector. If Hydrogen, Orchestrator (Carbon) and Orchestrator (Argon) are all absent, update. If only Hydrogen is absent, you are mid-rollout.

• Confirm the tier. The documentation table and the announcement post disagree about free access. Treat the paid tiers as the reliable floor.

• Confirm the setting. High is the default. Any comparison you read that quotes 54 is quoting Extra, and any comparison quoting 51 is quoting the model you get without touching anything.

• Confirm the string. gemini-4-hydrogen. No flash, no bare gemini-4, and no hyphen where a dot belongs.

• Confirm what runs the planning turn. Orchestrator's configuration is fixed and its availability depends on your plan, so the planner is not a variable you can tune — only one you can select or lose.

None of that is a reason to wait. It is a reason to test the thing you actually shipped rather than the thing the screenshot showed. The model is a day old, the rollout is still uneven, and the two published index figures for it differ by three points on a setting nobody mentions in the headline.