On Tuesday, 10 November 2026, Google opened Gemini 4 Carbon to paying developers, enterprises and Google AI Ultra subscribers. Two days later, on 12 November, it released Gemini 4 Hydrogen, a faster and cheaper model that is already rolling out in Search for every user worldwide. Taken together, the launches are less a single new Gemini than a deliberate split: one model built to spend tokens on hard, long-horizon work, and another built to run quickly, at scale, in the products most people actually touch. For anyone following the frontier Artificial Intelligence race in our Technology section, this is the moment Google’s two-speed strategy becomes explicit. It is also the moment to separate what the company shipped from what the benchmark charts imply.
Two launches, two audiences, one family
Carbon, Google says in a blog post on 10 November by Koray Kavukcuoglu, SVP at DeepMind and the company’s chief AI architect, is its strongest model yet. The company points to cybersecurity, multi-step software engineering and scientific reasoning as the areas where it wants Carbon judged. The public release follows a limited preview for cybersecurity defenders through Google’s Fairwind Program, which began about two weeks earlier. Gemini 4 Argon, launched only last month, “remains the ideal choice for the majority of software engineering and knowledge work,” Kavukcuoglu wrote. Carbon is reserved for jobs that “test the limits of a model’s intelligence,” and it is designed to use more tokens than Argon in order to explore, verify and loop.
Hydrogen is the other half of that design. In a follow-up post on 12 November, Google priced it at $1.50 per million input tokens and $7.50 per million output tokens, flat across a one-million-token context window. Carbon matches Argon’s introductory rate of $2 and $10, which doubles after 30 November 2026 to $4 and $20. As The Verge noted, even the higher rate is still, on a per-token basis, a fraction of the list prices Google cites for OpenAI’s GPT-6 Astra and Anthropic’s Fable 5.2. That comparison comes with an important caveat, which The Verge also made: token prices do not automatically translate into lower bills, because models consume tokens at different rates.
The more important split is access. Carbon is available through the Gemini API in Google AI Studio and Gemini for business for paying developers and enterprises, and to AI Ultra subscribers in Google Antigravity, the Gemini app and AI Mode in Search. It is not, at least for now, a free-tier model. Hydrogen is. It is rolling out across all tiers of AI Studio, the Gemini API, Gemini for business, the Gemini app, Google Search, and, for paid Pro and Ultra subscribers, Antigravity. Free users of the Gemini app get it through automatic model selection when Google decides it is the right fit.
Carbon: a frontier model that is not for everyone
The practical question for teams is not whether Carbon can fill a slide with benchmark wins. It is when the extra reasoning is worth the extra tokens, and who is allowed to use it. Google’s own framing is unusually clear on the second point. Carbon is a paid and Ultra product. If you are a free user of the Gemini app, or a Pro subscriber hoping the new top model will appear in your dropdown, the 10 November launch does not change your default. The Verge reported that it is not clear when, or if, Carbon will reach lower-tier subscribers.
To justify that restriction, Google is leaning on long-horizon agent work. In Google Antigravity, its agentic development platform, the company is previewing a “/swarm” command that turns a single prompt into a coordinated team of agents, visible in real time. Developers can assign Carbon to the roles that demand the deepest reasoning — architectural design, coordination, auditing — and pair it with faster models in isolated workspaces, including adversarial verifiers that test every change. The demos Google chose to publish are spectacular by design: a multi-agent reconstruction of its Platform 37 office from city planning PDFs, an open-world campus game, a playable historical fortepiano. They show what a well-prompted swarm can build overnight. They do not, by themselves, show how often that process finishes cleanly on an ordinary production codebase.
The more consequential claims sit in cybersecurity and science. Through Fairwind, “Carbon for cyber” is being offered to defenders as a system that can inspect software, trace how flaws chain across systems, and generate verified root-cause patches. Google says that in Project Fortify, its internal scanning, triage and remediation system, Carbon reached 100% recall on scanning and 95% precision on triage, on an internal benchmark, and that tests on open-source repositories showed more than a 25% relative improvement on Argon’s already strong recall. It also says Google’s Cloud Vulnerability Research team used Carbon to prove that two bugs in a popular open-source project were exploitable, at a severity it compared to the recent PHP2Shell case, and that Wiz, using Carbon in its Scan for Good program, found zero-day vulnerabilities in two major SaaS providers in under two hours. Those are striking results. They are also, for now, results reported by Google, with the open-source exploit details still under responsible disclosure.
Hydrogen: the fast model that now runs in Search
If Carbon is the model Google wants specialists to pay for, Hydrogen is the model it wants the rest of the internet to meet. Robby Stein, vice president of product at Google Search, told The Verge it is “the biggest model upgrade to Search yet,” and that it is “now the new default model in AI Mode” and is “being used for your hardest questions in AI Overviews.” TechCrunch, reporting the same launch on 12 November, described the Search pitch in plainer terms: Hydrogen is meant to handle messy, multi-part questions, hard-to-read typos and highly specific requests, and to answer each part of a query more precisely.
The Search changes Google is promising are familiar in kind, if not in degree. Hydrogen is said to improve query fan-out, the technique that breaks one prompt into several subqueries, and to return a wider set of links as a result. Responses in AI Mode are supposed to be more visual: comparison tables, graphs, maps, and, in Stein’s example, even an interactive driving simulation. Google also says language understanding has improved across all supported AI Mode languages, and that the update is rolling out to Search users globally. For a company whose core product still sits in front of billions of queries a day, this is the launch that will be felt first, and judged first, by people who will never open an API dashboard.
On the developer side, Hydrogen is framed as a major upgrade from Gemini 3.8 Flash. Google says that at its default high thinking setting it finishes tasks 2.5 times faster than 3.8 Flash, more than triples its Terminal-Bench 4.1 score (57% against 18.2%), and cuts fabricated claims by more than five times. It offers four thinking levels — low, medium, high (the default) and a new “xhigh” tier — so teams can trade latency and cost against depth. The surprising headline in Google’s own announcement is that Hydrogen, at that default high setting, scores 51 on the Artificial Analysis Intelligence Index, just above Argon’s 50, and reaches 54 at xhigh. If that result holds up under independent runs, it complicates the simple story that Carbon and Argon are the “smart” models and Hydrogen is merely the cheap one.
Orchestrator, Swarm, and the end of the single-model session
The strategic idea tying the two launches together is that a serious workflow should no longer belong to a single model from start to finish. Alongside Hydrogen, Google introduced Orchestrator in Antigravity: a mode that switches dynamically between models inside one session. In the example the company gives, Carbon starts in a read-only phase to explore a repository, resolve ambiguities and write a plan, then hands execution — edits, package updates, tests — to Hydrogen. If Hydrogen hits friction, control returns to Carbon. For parallel work, Carbon can act as coordinator while Hydrogen runs worker and verifier agents in isolated workspaces.
Google says that on Terminal-Bench 4.0, Orchestrator reached near-Carbon task success rates while, on average, using 24% less quota and finishing 19% faster. A companion post on the Antigravity blog reports the same percentages and describes two flavors: Orchestrator (Carbon) and Orchestrator (Argon) for AI Ultra subscribers, and Orchestrator (Argon) for paid Pro subscribers. That packaging is as revealing as the benchmark. Google is not asking most developers to become router experts. It is selling a default that hides the routing, and it is keeping the strongest coordinator behind the Ultra paywall.
This is a more honest picture of how frontier models are actually used in 2026 than the old habit of crowning a single “best” model. Planning, execution and verification have different cost-quality curves. The risk is the obvious one: a handoff can drop context, a fast worker can implement a flawed plan efficiently, and a coordinator can spend its budget re-planning. Google’s numbers say the trade is worth it, on the benchmarks it chose. Independent evaluations of Orchestrator, on tasks the vendor did not pick, have not yet had time to accumulate. They will matter more than the launch-day percentages.
What the benchmarks show, and what they leave open
Google arrived this week with a long list of first-place finishes. Carbon, it says, ranks first on CursorBench 4.0, on Perplexity’s WANDR and on Zapier’s AutomationBench, at more than two times lower cost per task than the next-best external model, according to Google. On the new Multimodal Knowledge Agent benchmark, built with Artificial Analysis to test agents over dense files that mix charts, drawings and tables, Carbon ranks second, at more than six times lower average cost per task than the top model. Hydrogen, for its part, is said to lead its class on several professional-agent tests, to rank fifth on Agent Arena, and to match or beat competing models in its tier, including, in Google’s comparison as reported by The Verge, Claude Sonnet 5.5 and GPT-6.1 Terra.
Those figures should be read with the usual discipline. Vendors select the suites they publish. Many of the numbers above come from Google’s own posts, not from a third-party leaderboard run after the models were in general availability. The Verge, in its Carbon report, put the point bluntly: companies can and do cherry-pick benchmark results, and Google’s charts place Carbon among the most powerful models on the market, behind Anthropic’s Fable 5.2, which launched on 9 November. A one-point edge on a composite index, or a first-place finish on a brand-new vendor-partnered test, is a starting claim, not a settled ranking.
What is harder to dismiss is the product structure. Google now has a fast model it is willing to put in front of every Search user, a mid-frontier workhorse in Argon, a top-end reasoner in Carbon, and a router that tries to combine them. The competitive question is no longer only “whose single model scores highest.” It is whether Google can make that stack cheaper, more reliable and more present in daily software than the stacks at Anthropic and OpenAI. Pricing through 30 November is clearly meant to encourage teams to find out. After that date, Carbon’s list price doubles, and the economic argument will have to stand without the introductory discount.
Safety claims, and what to watch next
Because Carbon is being sold in part as a cybersecurity model, Google’s safety section is longer than a routine launch note. The company says Carbon blocked 99% of attack trajectories in a simulated environment seeded with ExploitBench and CyScenarioBench, cutting the per-attempt breach rate to 0.8% from 3.1% on Argon. It reports that Trajectory Labs found no new universal jailbreaks across 132 hours and more than 13,000 requests. Most eye-catching is the claim, first shared when Carbon was previewed, that the model reached a 0% attack-success rate on the Gray Swan indirect prompt-injection benchmark — “the first time that milestone has been reached,” Google wrote. In the same breath, the company added a necessary qualification: automated red-teaming has already shown that models, including this one, can still be improved against sophisticated attacks in some distributions. A zero on one benchmark is not a proof of immunity.
Hydrogen inherits the Gemini 4 safety stack and adds numbers aimed at everyday agent use: Google says it reduces overreach in agentic tasks by up to 81% and under-service by up to 79% against 3.8 Flash, and that it is more than 2.3 times more resilient on the same Gray Swan prompt-injection test. For a model that will sit inside Search and the Gemini app, those claims will be tested in public, by adversaries who do not need a Fairwind invitation.
The near-term watch list is concrete. First, whether independent evaluations, once they catch up, still show Hydrogen punching above its price and Carbon justifying its token appetite. Second, whether Orchestrator and /swarm survive contact with messy repositories, or remain launch-day demonstrations. Third, how Search actually feels with Hydrogen as the default in AI Mode: better links and clearer answers, or a more confident overview that still needs a careful human check. Fourth, the 30 November price reset, which will tell us how much of Carbon’s early adoption was a sale. And fifth, whether Argon really does remain the daily workhorse, or whether teams simply point everything at Hydrogen and call Carbon only when something breaks.
Google has not, in these two days, crowned a single successor to its last frontier model. It has shipped a family with a division of labor, and it has put the cheapest, fastest member of that family into Search. That is a more interesting move than another isolated benchmark dump. It will take longer than a week to know whether the division of labor holds.
