Google Search Adds Gemini 4 Hydrogen as Orchestrator and Decision 1 Ship
On 11 and 12 November 2026 Google posted Gemini 4 Hydrogen for Search, Orchestrator for AI Ultra subscribers, and a Decisions API on Gemini Decision 1. The notes name products and audiences. They do not attach a shared spec.
Google posted three product notes across 11 and 12 November 2026, and a fourth note from DeepMind that lists who can reach a rollout without naming the model in the text. Taken as a set, the written record is specific about names and audiences and thin about everything a buyer would put in a comparison sheet. Google’s Search post says Search is getting its biggest model update yet with Gemini 4 Hydrogen, rolling out globally for all users that day. Google’s Decisions post, from the evening before, introduces a Gemini Decisions API in Google AI Studio, powered by a model it calls Gemini Decision 1. A later post the same UTC day as the Search note says Orchestrator is rolling out inside Antigravity for Google AI Ultra subscribers. DeepMind’s note, earlier that afternoon, maps developer, enterprise, and general access and then stops at an attached photo.
None of those sentences is a benchmark. None of them is a price. None of them is a changelog with versioned endpoints. A team that treats the cluster as proof that one model now sits inside Search, the API, and a multi-model workspace is reading past the page. The posts do not say that. They sit near each other in time. Proximity is not a shared specification.
What the Search post actually commits to
The Search post is time-stamped 12 November 2026 at 19:23 UTC. The claim in the first sentence is large and unqualified: Google Search is getting its biggest model update yet, the model name is Gemini 4 Hydrogen, and the rollout is global, for all users, that day. “Biggest” is Google’s ranking of its own history. The post does not show the prior update it is measuring against, and it does not say whether “biggest” means quality, context, traffic coverage, or the size of the engineering change. A reader can accept the sentence as a product announcement and still refuse it as a measurement.
The only capability the written text names is intent. Search, under this update, “better understands intent.” The example that follows is a usage description, not a scored task: ask nuanced, multi-part questions in a natural flow of thought, even with specific requests. The sentence breaks off there. The link on the post resolves to a video attachment on the same status, not to a separate article. This piece uses the written text. It does not treat unseen frames as a source, so the list of “what’s new in Search” stops where the text stops.
That limit matters. A multi-part question is a familiar failure mode for older search boxes: the system answers the first clause, drops a constraint, or returns a page that matches keywords rather than the relation between clauses. Google is saying Hydrogen is meant to hold that relation. It is not saying how often it succeeds, in which languages, on which classes of query, or whether the change is on by default for signed-out users in every country the word “globally” might suggest. “All users” and “globally” are the post’s words. They are not accompanied, in the text, by an exception list. Absence of an exception list is not the same thing as a published coverage map.
For a publication that covers tools rather than slogans, the useful residue is small and solid. On 12 November 2026 Google attached the name Gemini 4 Hydrogen to Search, described the change as a model update rather than a feature toggle with a separate brand, and pointed the user-facing benefit at multi-part questions asked in ordinary language. Anyone briefing a search team can repeat those three points and then stop. The next sentence in a slide deck, if it adds a percentage or a latency figure, is no longer the post.
A video is not a spec sheet
Official accounts often put the detail in an attachment and the headline in the text. That pattern is legible here. The Search post tells a reader that something large shipped, then points at a video on the same status. A video can show a query, a result, and a voiceover. It cannot be cited the way a changelog can, unless someone transcribes it and the transcription is the source. This article does not do that transcription. The practical effect is that Hydrogen’s Search behavior, beyond intent and multi-part questions, is not in the written record used here.
That is a narrower claim than “the update is empty.” Google may have shown more in the video. A team that needs the more, for a vendor review or an internal memo, has to watch the attachment itself and write down what it shows. Until that happens, the honest summary is the text: a named model, a global rollout claimed for that day, and one described improvement.
There is also a naming point worth keeping straight. Gemini 4 Hydrogen is the name in this post. The post does not say Hydrogen replaces every other Gemini surface, and it does not say the Search update is the same binary as the API model a developer would call the next morning. Search and the Gemini API are different products even when a company uses one family name for both. This post talks about Search.
Orchestrator is a session design, not a Search feature
About an hour and a half later, at 20:50 UTC on the same day, Google posted a different note. Orchestrator is rolling out in Antigravity, and the audience in that sentence is Google AI Ultra subscribers. The behavior it describes is not a better answer to a search query. It is a change in how a session is routed. Instead of sending the entire task to one model, a subscriber can run multiple models in a single session.
The next sentence is the design claim, and it is more concrete than the Search post’s intent line. Orchestrator pairs a frontier model with a fast execution model. Carbon sets the plan, then “hands off.” The written text ends on that verb. The fast execution model is not named in the fetched wording. Carbon is named, and it is placed in the planning role. The post does not give Carbon a version number, a context length, or a benchmark. It does not say Carbon is Hydrogen, and it does not say Carbon is a separate family. A reader who merges the two names because they appeared on the same UTC date is doing work the posts do not do.
The Ultra restriction is easy to skip if someone is collecting model names and ignoring audiences. The Search post says all users. The Orchestrator post says Ultra subscribers. Those are different doors. A person who can type a multi-part question into Search on 12 November is not, on the evidence of these posts, the same person who can open an Orchestrator session. Nothing in the Search text mentions Antigravity. Nothing in the Orchestrator text mentions Search or Hydrogen.
What the Orchestrator text does support is a product argument about split roles. A single-model session has one context, one style of failure, and one bill for the whole task, whatever the pricing page later says. A paired session, as described, puts planning on one model and execution on another, with a handoff between them. That can be a way to spend a slower, more capable model on the plan and a faster model on the steps. It can also be a way to create a new failure at the boundary: the plan is coherent and the execution model drops a constraint, or the execution model is fast and the plan was wrong. The post does not discuss that boundary. It announces the pairing. The risk at the handoff is an inference from the shape of the feature, not a finding Google published in the text.
For tool selection, the sentence that should survive a meeting is this: Orchestrator, inside Antigravity, for Ultra subscribers, is a multi-model session in which Carbon sets the plan and a fast execution model, unnamed in the text, is the other half of the pair. If a vendor questionnaire asks which model executes, the correct answer from this post is that the post does not say.
Decision 1 is a different job from generation
The third Google post is earlier, and it should stay earlier in the narrative. At 18:26 UTC on 11 November 2026, Google introduced the Gemini Decisions API in Google AI Studio, powered by a new model it names Gemini Decision 1. The contrast the post draws is the whole substance of the written claim. Instead of generating a text response token by token, Decision 1 evaluates an input against up to 1,000 questions. The inputs it lists are text, images, audio, video, and PDFs. The sentence continues past that number, into a link this article did not open, so the written record stops at the evaluation claim and the input types.
That is already enough to separate Decision 1 from Hydrogen-in-Search and from Orchestrator. Search, in the later post, is a place a person asks a question and expects an answer shaped for a results page. Orchestrator is a session that plans and hands off. Decision 1, in its own post, is an evaluator: many questions, one input, not a token-by-token text reply. A developer who needs a paragraph, a summary, or a drafted email is not looking at the job this post describes. A developer who needs a batch of yes-or-no or scored checks against a document, an image, a clip, or a PDF is closer to the wording, with the caveat that the post does not publish the response schema. “Evaluates your input against up to 1,000 questions” does not, by itself, say whether each question returns a boolean, a score, a label, or a short string. The cap of 1,000 is in the text. The shape of each answer is not, in the portion fetched.
The 1,000 figure is the only quantity in the four posts that behaves like a product limit rather than a date or a clock time. It should be quoted with the verb attached. Decision 1 evaluates against up to 1,000 questions. It is not a promise that every call will ask 1,000 questions, and it is not a rate limit, a price tier, or a latency budget. Those readings are available to anyone who wants a rounder story. They are not in the post.
Placing the API in Google AI Studio tells a developer where Google wants the first calls to happen. It does not say the API is absent from other surfaces. It also does not say Decision 1 is the model behind the Gemini API generally. The name is specific: Gemini Decision 1, inside a product called the Gemini Decisions API. A team that writes “Google replaced generation with decision calls” has expanded one product into the whole platform. The post does not make that expansion.
There is a practical reading that stays inside the wording. Some production systems do not want a fluent paragraph. They want a stable check: does this PDF mention a required clause, does this image contain a logo, does this audio clip match a listed claim. Token-by-token generation is an awkward tool for that job, because the model can explain, hedge, or drift in format. An API that evaluates an input against a list of questions is aimed at the check, not the explanation. Whether Decision 1 is reliable at that aim is not something these posts measure. The posts only show that Google is willing to ship a named model whose described behavior is evaluation at that scale of questions, across those media types.
DeepMind lists doors, and does not name the model
At 18:02 UTC on 12 November, about an hour and twenty minutes before the Search post, Google DeepMind posted an access list. The text begins “Now rolling out for” and then splits the audience.
Developers are told to build with the Gemini API in Google AI Studio and in Android Studio, and in Antigravity for paid Google AI Pro and Ultra subscribers. Enterprises are told to deploy via Gemini for business and via Antigravity. Everyone is told to access “it” in Google Search, and the sentence ends at a photo attachment on the same status. The photo is not a source for this article. The written text does not contain the names Hydrogen, Carbon, Decision 1, or Orchestrator.
That missing name is the most important fact in the DeepMind post. The pronoun “it” has no antecedent in the fetched wording. A reader who already saw the Search post can guess that “it” is Hydrogen, because both posts mention Google Search on the same UTC date and the Search post names the model. That guess is a reading across two accounts. It is not a sentence DeepMind wrote in the text used here. A careful note should say so. Otherwise a later summary will cite DeepMind for a model name DeepMind’s text did not use.
The access split is still usable, and it does not need the missing name to be useful. Three doors are described.
The developer door has two rooms. The Gemini API in Google AI Studio and Android Studio is listed without a paid-tier adjective in that clause. Antigravity, in the same developer sentence, is limited to paid Google AI Pro and Ultra subscribers. A developer with an unpaid AI Studio project is not, on this wording, the same as a developer with Antigravity access. The post does not say the API and Antigravity expose the same model.
The enterprise door names two deployment paths: Gemini for business, and Antigravity. It does not describe a contract, a region, a data-residency option, or a support tier. “Deploy via” is the verb. What gets deployed is the same unnamed “it.”
The general door is Search, for everyone, which lines up in audience with the Search post’s “all users,” even though only the Search post names Hydrogen. The alignment is about who can reach Search, not about a shared technical appendix.
Put the Ultra line from the Orchestrator post next to the Pro-and-Ultra line from DeepMind and the difference is sharp. DeepMind says Antigravity is available to paid Pro and Ultra subscribers. Google’s Orchestrator post says Orchestrator, inside Antigravity, is rolling out to Ultra subscribers. The texts do not say Pro subscribers receive Orchestrator. They also do not say Pro subscribers are excluded from every Antigravity feature. The safe statement is the narrower one: Orchestrator is described as an Ultra rollout, and Antigravity access is described, in a different post, as Pro and Ultra. A procurement note that collapses those into “paid subscribers get the multi-model session” is not supported.
Four posts, three product names, no shared table
A reader who wants one story will be tempted to build it. Hydrogen is the Search model. Carbon plans inside Orchestrator. Decision 1 evaluates inputs against long question lists. DeepMind tells developers, enterprises, and everyone where to stand. The dates fit in a window of about twenty-six hours, from the Decisions post on 11 November to the Orchestrator post on 12 November. A launch cluster is a fair description of the timing. A single platform release, with one model card, is not what the texts provide.
The names that do appear should stay attached to the sentences that carry them.
Gemini 4 Hydrogen appears in the Search post, tied to a global Search rollout for all users on 12 November, with intent and multi-part questions as the described change. It does not appear in the Orchestrator text or the Decisions text. It does not appear in the DeepMind text used here.
Carbon appears in the Orchestrator post, as the model that sets the plan, inside a pairing of a frontier model and a fast execution model, for Ultra subscribers in Antigravity. It does not appear in the Search post.
Gemini Decision 1 appears in the Decisions post, as the model behind the Gemini Decisions API in Google AI Studio, evaluating inputs across five media types against up to 1,000 questions, in contrast to token-by-token text generation. It does not appear in the Search post.
Antigravity appears in the Orchestrator post and in the DeepMind post. In one, it is the home of Orchestrator for Ultra. In the other, it is a build and deploy surface for paid Pro and Ultra subscribers, and an enterprise path next to Gemini for business. The two mentions can both be true. They are not the same feature list.
Google AI Studio appears twice, in different jobs. The Decisions API is introduced there. Developers are also told to build with the Gemini API there. An API for evaluation and an API for building with Gemini are not interchangeable just because the studio name matches. A developer reading both posts still has to check, in the studio itself, which model picker corresponds to which post.
Android Studio appears once, as a developer build surface for the Gemini API, in the DeepMind list. No feature inside Android Studio is described.
Gemini for business appears once, as an enterprise deploy path. No seat count, no admin control, and no model name ride along with that phrase in the fetched text.
This is the inventory. It is shorter than a launch blog usually is, and that shortness is the story. Google’s own accounts chose to announce these surfaces in posts whose written lines name products and audiences and then point at media. The media may carry demos. The written record, which is what can be quoted without inventing a transcript, does not carry a comparison table, a price, or a score.
What a developer can do with this record
A developer who has to decide where to spend a day can still extract a sequence of checks from the posts, without pretending the posts finished the evaluation.
First, Search. If the question is whether a public, signed-in or signed-out user might already be hitting Hydrogen when they search, the Search post says the update was rolling out globally for all users on 12 November. That is a reason to look at live Search behavior on multi-part queries, and to write down what the live product does. It is not a reason to cite a quality gain the post did not measure. “Better understands intent” is the claim to test, not a result to paste into a report.
Second, Google AI Studio. Two different invitations point there. One is the Decisions API and Decision 1. The other is the Gemini API for developers, in the DeepMind list. A developer can open the studio and see which of those is actually offered in the account they have. The posts do not list model IDs, so the studio screen is the next source, not a number remembered from a headline.
Third, Android Studio, if the work is on Android. The DeepMind post names it as a place to build with the Gemini API. It does not describe a new Android API, a minimum IDE version, or a device requirement. The check is whether the IDE build the team already uses shows the surface the post is pointing at.
Fourth, Antigravity, and only with the tier in view. Pro and Ultra are the paid subscribers DeepMind associates with Antigravity. Ultra is the tier Google associates with Orchestrator. A Pro account that can open Antigravity has not, on these texts, been promised Orchestrator. An Ultra account has been told Orchestrator is rolling out, with Carbon in the planning role and an unnamed fast model on the other side of the handoff. Trying the session is a product check. Quoting a speedup is not, because no speedup figure is in the text.
Fifth, the Decisions API, if the job is evaluation rather than drafting. The inputs the post lists are text, images, audio, video, and PDFs. The question cap is up to 1,000. A team with a review workflow, a compliance checklist, or a media screening step has a named API to inspect. A team that needs long-form generation should not reroute that need into Decision 1 on the strength of a shared Gemini prefix.
None of those checks requires a claim about which model is “best.” The posts do not rank Hydrogen, Carbon, and Decision 1 against each other. They do not even put the three names in one paragraph.
What an enterprise buyer cannot conclude
Enterprise language shows up in one place: DeepMind’s line that enterprises deploy via Gemini for business and via Antigravity. That is an invitation to a path. It is not a case study. No client is named. No deployment time is given. No statement says an existing contract now includes Hydrogen, Carbon, or Decision 1.
A buyer who needs residency, retention, admin audit, or a private endpoint will not find those words in these posts. Adding them because “enterprise” usually implies them is how vendor notes drift. Gemini for business may have those controls on a page this cluster does not link in the fetched text. Until that page is the source, the buyer note should say the posts name the path and do not describe the controls.
The same restraint applies to cost. Ultra, Pro, and “paid” tell a buyer that some doors are not free. They do not say the monthly price, the overage, or whether Decision 1 calls are billed as input tokens, output tokens, or something else. The Decisions post contrasts Decision 1 with token-by-token generation. That contrast is about how the model responds, not about the invoice. A finance partner who hears “not token by token” and concludes “not billed” has changed the subject. Billing is not discussed in the fetched wording.
Support and rollout risk are also outside the text. “Rolling out” can mean a flag is on for every account, or a flag is moving across accounts through the day. The Search post says globally, for all users, today, which is the strongest coverage sentence in the set. The Orchestrator post says rolling out to Ultra subscribers, which can be read as a start rather than a finished switch. DeepMind says “now rolling out,” which is the same family of phrase. A status page would settle the difference. These posts are not a status page.
There is a governance point that does follow from the audience split, without extra facts. A company that allows employees to use consumer Search is, if the Search post is accurate, exposing those employees to Hydrogen as the search model. A company that allows only a managed enterprise SKU is pointed at Gemini for business and Antigravity, with the model unnamed in that post. The two policies are not automatically the same model, the same log, or the same retention. Security and legal teams should not assume a consumer Search update and an enterprise deploy path are one control plane because the same week produced both announcements.
How to brief this without inflating it
The useful brief fits on a single page, and it should, because the sources fit in four short posts.
On 11 November 2026, Google introduced the Gemini Decisions API in Google AI Studio. The model name is Gemini Decision 1. It evaluates an input, which may be text, an image, audio, video, or a PDF, against up to 1,000 questions, instead of generating a text response token by token. The rest of that sentence is not in the written excerpt.
On 12 November 2026, DeepMind posted an access list for an unnamed “it”: Gemini API in Google AI Studio and Android Studio for developers; Antigravity for paid Pro and Ultra; Gemini for business and Antigravity for enterprises; Google Search for everyone. A photo is attached. The model name is not in the text.
Later that day, Google said Search’s biggest model update yet is Gemini 4 Hydrogen, rolling out globally for all users, with better intent understanding for nuanced, multi-part questions. A video is attached. No score is in the text.
Later still, Google said Orchestrator is rolling out in Antigravity to AI Ultra subscribers, so a session can use more than one model. Carbon sets the plan. A fast execution model is the other half of the pair and is not named in the text.
That brief is the whole factual core. Analysis can sit under it, and it should stay labeled as analysis. The split between a public Search model and a paid multi-model session is a real product split, because the audiences differ. The split between a generative reply and an evaluation against a question list is a real job split, because the Decisions post draws it. The refusal to treat DeepMind’s “it” as Hydrogen is a sourcing split, not a claim that the photo shows something else. The photo was not used.
What should not sit under the brief is a ranking, a savings estimate, or a sentence that Hydrogen, Carbon, and Decision 1 are three modes of one release. The posts would have been easy to write that way if that was the message. They were not written that way. Each name arrives in its own post, with its own verb: Search gets Hydrogen, Orchestrator pairs Carbon with a fast model, Decision 1 evaluates.
Questions the posts leave open
A follow-up that would turn this cluster into a tool review is easy to list, and listing it is fair as long as the list is not disguised as findings.
Which model ID in Google AI Studio corresponds to Decision 1, and which corresponds to whatever the DeepMind post meant by the Gemini API? The posts do not say.
What is the fast execution model in Orchestrator, and is Carbon the frontier model in the pair or only the planner? The colon in the post puts Carbon on the plan. It does not print a second name.
Does “all users” in Search include signed-out traffic, every language the product supports, and every country where Search operates? The post says globally and all users. It does not print the matrix.
Does an enterprise deployment through Gemini for business include Hydrogen, Decision 1, both, or neither? The enterprise sentence does not name a model.
Is Orchestrator available the moment an Ultra subscription is active, or is the rollout still moving through that tier? The post says rolling out. It does not say finished.
What does a Decision 1 response look like when a question is ambiguous, when the input is a long PDF, or when the question list approaches 1,000? The cap is published. The failure mode is not.
Those questions are the difference between an announcement and a review. This article is the first. It can tell a reader what Google and DeepMind committed to in writing on those two dates. It cannot score the products, because the posts do not score them, and no separate measurement is in the record used here.
Why the cluster still changes the tool map
Thin posts can still move a map if the names and the doors are new relative to the prior week of public notes from the same accounts. Hydrogen is a Search model name a team will now hear in user complaints and in “why did this query change” tickets. Carbon is a planning name a team will now hear inside a paid workspace, if the Ultra rollout reaches their account. Decision 1 is an API name for people who have been forcing checklists through chat completions and fighting the format. Antigravity is confirmed, in this window, as a place Google is willing to put a multi-model session and as a place DeepMind lists for paid subscribers and enterprises.
The map change is about where to look, not about who won. Search is the wide door, described as all users, with a named model and a narrow written benefit. Antigravity is the paid door, with a session design that splits plan and execution, and with a tier detail that Pro and Ultra do not automatically share. Google AI Studio is the developer door, hosting both a general build invitation and a specialized decisions API. Gemini for business is the enterprise phrase, present and undescribed.
A team that updates its internal tool list this week can add those four lines and a date. It should not add a quality delta, a price, or a sentence that one subscription now covers every name in the cluster. The posts are the source. They are also the boundary.
Primary records for the two claims that carry product names in Google’s own wording are the Search post on Gemini 4 Hydrogen and the Decisions API post on Gemini Decision 1. The Orchestrator wording and the DeepMind access list cited above come from Google’s post at 20:50 UTC on 12 November 2026 and DeepMind’s post at 18:02 UTC the same day. Where a sentence in those posts ran into an attachment, the attachment was not treated as text.