Gemini 4 Carbon is out: Google's strongest model opens to every paying developer — and its launch price doubles on December 1
Google put its strongest model on general sale on November 10. Gemini 4 Carbon, the frontier system the company had kept behind a cybersecurity gate since October, is now available to every developer and enterprise paying for the Gemini API, and it is rolling out to Google AI Ultra subscribers. It arrives with a date attached: the introductory price of $2 per million input tokens and $10 per million output tokens runs through November 30 — from December 1, the same tokens cost exactly twice as much.

Artwork: Google (official blog) — the official launch key art for Gemini 4 Carbon, published by Google with the announcement. Carbon is a software model; there is no machine to photograph.
The strongest model Google has, now on sale
Carbon is the top of Google's Gemini 4 line: the model the company built, in its own words, "for tasks that test the limits of a model's intelligence" — complex, multi-step coding and debugging, open-ended scientific research, and long-horizon agent work that runs for hours rather than turns. It is built on Gemini 4 Argon, the frontier model Google released in October, and the difference between the two is mostly a matter of appetite. Carbon spends more tokens on a task than Argon does: it explores more paths, reasons longer, and runs what Google calls extended verification loops — checking its own work before it hands the work back.
The engineering numbers are on Google's model documentation. Carbon works with a single unified budget of 1,048,576 tokens — roughly a million — that covers the prompt, the thinking and the answer together, and it supports long decodes that let a single generation run on toward that same million-token ceiling instead of stopping at a short output limit. Its thinking comes in three levels — low, medium and high — and its knowledge cutoff is March 2026. It reads text, images, audio and video; it answers in text. Google is explicit about the division of labor inside its own family: Argon "remains the ideal choice for the majority of software engineering and knowledge work," and Carbon is the escalation — the model a system calls when the job is the hardest one on the board.
One family, three elements, five weeks
Carbon did not arrive alone, and the calendar matters. Google opened the Gemini 4 family on October 13 with Argon, its efficient frontier model. On October 28 it announced Carbon — but handed it only to a set of cybersecurity defenders through its Fairwind Program, saying it wanted feedback and time on the guardrails before a wider release. On November 10, Carbon went general. And on November 12, two days later, Google introduced the third element: Gemini 4 Hydrogen, the fast, efficiency-first model that carries Gemini 4 to everyone — all tiers of AI Studio and the API, the Gemini app, Search, and paid Pro and Ultra subscribers in Antigravity — at a flat $1.50 per million input tokens and $7.50 per million output tokens across its full million-token window.
Three frontier-class releases in five weeks is a cadence the industry has not seen from Google before, and the naming is the org chart: Argon for the everyday frontier workload, Carbon for the hardest problems, Hydrogen for speed at population scale. Hydrogen even beats Argon on the Artificial Analysis Intelligence Index at its default thinking level, Google says — 51 to 50, and 54 at its new extra-high setting. But the strongest of the three is Carbon, and Carbon is the one behind the paywall twice over.
Where it runs — and who cannot have it
Developers reach Carbon through the paid tier of the Gemini API, in Google AI Studio and Android Studio, under the model ID gemini-4-carbon. Enterprises reach it through Gemini for business, through the API, in Google Antigravity, or in CodeMender. Consumers reach it in exactly one way: a Google AI Ultra subscription, which unlocks Carbon in the Gemini app, in AI Mode in Search and in Antigravity. There is no free tier — Google's pricing page states it in a note under the model's name — and the AI Pro and AI Plus plans that most paying Gemini users actually hold do not include it in the app. Scientists get a side door: Google expanded its research program — free top-tier Antigravity access for six months, originally opened for Argon to 10,000 scientists — to include Carbon for current and future participants.
One part of Carbon is not on sale at any price. The specialized cybersecurity variant — model ID gemini-4-carbon-cyber-permissive — is generally available on the Gemini Enterprise Agent Platform only behind an allowlist, through the Fairwind Program. Google's documentation draws the line plainly: the public model is for everyone who pays; the cyber model is for vetted defenders who apply.
/swarm: one prompt, a team of agents
The launch's second act is a preview inside Google Antigravity, the company's agent-first development environment. A command called /swarm turns a single prompt into a coordinated team of agents that a developer can track in real time. Carbon takes the roles that demand the deepest reasoning — architectural design, coordination, auditing — while faster, lighter models work in parallel across isolated workspaces, and adversarial verifier agents test every change the team produces. It is the clearest statement yet of where Google thinks the frontier model earns its price: not answering one hard question, but supervising a long campaign of them. Carbon's engineering profile points the same way — multi-package migrations, root-cause investigations that cross subsystems, architectural redesigns, and web development loops in which the model builds, runs, recovers from its own errors and inspects the rendered interface, not just the code.
The scoreboard, with the asterisks printed
Google's claims for Carbon are specific, and they are Google's. The company says Carbon ranks first on CursorBench 4.0, on Perplexity's WANDR and on Zapier's AutomationBench — three benchmarks built around real agentic work — at more than twice lower cost per completed task than the next best model from outside Google. On the new Multimodal Knowledge Agent benchmark, built with Artificial Analysis to test agents reasoning over dense files full of charts, images, blueprints and tables, Carbon ranks second while costing more than six times less per task than the model in first place. In finance, Google says Carbon matched the best model tested on Hebbia's end-to-end financial agent evaluation and finished first among 23 frontier configurations on Balyasny Asset Management's general financial intelligence benchmark.
The independent scoreboard is less tidy, and this magazine prints it in the same size. The Artificial Analysis Intelligence Index — the ranking Google itself cites when Hydrogen beats Argon — puts Carbon at high effort on 60 points, behind Anthropic's Claude Fable 5.2 at its maximum setting on 65, as the index stood in the days after the launch. Google's counter-argument is economic rather than absolute: on the workloads it chose to publish, Carbon finishes the task at half the cost or less. Both things can be true at once, and for the moment both are: the strongest model Google has ever shipped is not, by the independent index, the strongest model on the market — it is the one arguing it is close enough, at a lower price per finished job. That argument gets more expensive to make on December 1.
The cyber model stays behind a gate
Cybersecurity is where Carbon was born — it spent its first two weeks of life exclusively with defenders — and the results Google reports from that world are the most concrete in the launch. Accessed through the Fairwind Program, the company says, Carbon leads on the finding, exploiting and patching benchmarks it names: SEC-bench Pro, ExploitGym and CWE-bench v1. Inside Google's own Project Fortify — the autonomous system that scans, triages and remediates code — Carbon reached 100% recall on scanning and 95% precision on triage on Google's internal benchmark, and improved recall on open-source repositories by more than 25% over Argon's already strong rate, surfacing real logic flaws that had gone undetected. Google's Cloud Vulnerability Research team used it to prove that two bugs in a popular open-source project were in fact exploitable vulnerabilities, comparable in severity to the recent PHP2Shell exploit; the maintainers reclassified the bugs as critical and shipped a patch. And in Wiz's Scan for Good program, which assesses vulnerabilities in public digital infrastructure, Carbon found zero-day vulnerabilities in two major SaaS providers in under two hours — work Google describes as normally taking weeks.
The defensive framing is doing real work here. A model that can find and prove exploitable flaws is, by construction, a model that understands offense; Google's answer is distribution, not denial. The cyber capabilities ride in the separate, allowlisted variant, vetted through Fairwind, while the public Carbon — the one any paying developer can call today — carries the family's safeguards without the specialized cyber build. It is the same two-key logic the nuclear industry would recognize: capability in one room, permission in another.
The safety numbers Google wants on the record
Because Carbon is stronger where it is dangerous, Google published an unusually specific safety ledger with the launch. In simulated cyberattack environments seeded with its ExploitBench and CyScenarioBench tests, Carbon's safeguards blocked 99% of the attack trajectories flown by the company's adaptive red-teaming agent, and cut the per-attempt breach rate to 0.8%, down from 3.1% on Argon. In external testing by Gray Swan, the automated cyberattack query success rate fell to 2.17%, against 9.86% for Argon. Trajectory Labs, testing externally, found zero new universal jailbreaks across 132 hours and more than 13,000 requests. And on Gray Swan's indirect prompt-injection benchmark, Carbon recorded a 0% attack success rate — the first time any model has reached that mark, Google says, while adding in the same breath that prompt-injection robustness "remains an area for continued improvement."
The behavioral numbers point the same direction. On Google's 375-task agentic honesty benchmark, Carbon passed 92% of tasks, against 70% for Argon: it declined to fill in missing experimental data 98% of the time, and flagged methodological flaws in the work it was given 68% of the time. For a model being sold as an autonomous worker — one that will run for hours inside a company's systems — the claim that matters is not that it is brilliant. It is that it stops, says the data is not there, and refuses to invent it. Those are Google's measurements, run on Google's tests; the magazine notes the provenance because the industry has learned to. But they are, at least, numbers — published, falsifiable, and now on the record for competitors to attack.
Science at machine speed
The launch's most striking passages are not about code at all. Google's researchers describe Carbon working inside Antigravity on three scientific fronts. In formal verification, agents built a framework called AlphaProof Code and used Carbon to translate the security goals of the Signal protocol's cryptographic code into mathematical theorems, then generated millions of lines of Lean proof code — proving key security properties in weeks, a task Google says historically takes expert teams years. In a working wet lab's data, Carbon went past standard analysis of enzyme assay notebooks and caught the artifacts humans had missed: incomplete automated mixing in deep wells, chromatography peaks that needed deconvolution, trace sample carryover in the instruments. And in protein structure prediction — where models starve for evolutionary relatives of obscure proteins — Carbon agents searched petabase-scale unannotated sequence archives and, across seven low-homolog protein families from the CASP17 blind test, increased the homologs retrieved forty-fold on average. Axiom Bio, a drug-toxicity research company, reports Carbon is now its leading model for translational science, finding toxicity mechanisms it had not found before, and is moving production analyses onto Carbon and Argon together.
None of this is peer-reviewed, and all of it is reported by the vendor. But the shape of the claim is new. The argument for the frontier model used to be that it answers harder questions. Google's argument for Carbon is that it does longer work — campaigns of research with verification loops inside them — and the science stories are chosen to prove duration, not cleverness.
What thrills
- Duration as the frontier: a unified million-token budget and long decodes built for work measured in hours, not turns.
- A published safety ledger: 99% of simulated attack trajectories blocked, 0% prompt-injection success — numbers rivals can now shoot at.
- Cost per finished task: first place on three real-work benchmarks at over 2x lower cost per task, on Google's accounting.
What to watch
- The December 1 doubling: every rate on the card doubles when the introductory window closes — the launch price is a trial, not a price.
- Vendor benchmarks: the independent index still puts Anthropic's Fable 5.2 ahead; Carbon's case is economic, not absolute.
- Two paywalls: no free tier, and consumers need the top Ultra plan — Pro and Plus subscribers watch from outside.
A frontier with an expiry date
Strip the launch to its mechanics and three facts remain. The first is cadence: Argon on October 13, Carbon announced on October 28 and opened on November 10, Hydrogen on November 12 — a full frontier family, at three speeds, in five weeks, from a company that used to ship one flagship at a time. The second is structure: the strongest model is sold like a professional instrument — paid API, enterprise platform, top-tier subscription — while the model given to everyone is the fast one, a division that tells you exactly where Google believes the money and the risk both live. The third is the date. An introductory price with a published expiry, doubling on a named day, converts the launch into a trial window with a deadline: every team experimenting with Carbon this month is doing so at half the price its December budget will pay, on a model that — by Google's own design — spends more tokens per task than the one it replaces.
What Google has actually shipped, underneath the element names, is a claim about how frontier AI will be bought: not by the token, but by the finished task, verified. Carbon is the test of that claim — the strongest model the company has, the most specific safety ledger it has ever published, and a price that doubles in eighteen days. The market's answer arrives on December 1, when the trial ends and the meter starts running at full rate.
𝐆𝐄𝐓 𝐅𝐑𝐄𝐄 𝐅𝐔𝐋𝐋 𝐌𝐀𝐆𝐀𝐙𝐈𝐍𝐄 𝐀𝐓 thebuyerspulse.com

The data sheet
Gemini 4 Carbon (Google)
| Field | Detail |
| Model ID | gemini-4-carbon (public access); gemini-4-carbon-cyber-permissive (cyber variant, allowlisted) |
| Announced | October 28, 2026 — limited to cybersecurity defenders via the Fairwind Program |
| General availability | November 10, 2026 — paid Gemini API (AI Studio, Android Studio), Gemini for business; rolling out to Google AI Ultra subscribers (Antigravity, Gemini app, AI Mode in Search) |
| Context | Unified 1,048,576-token budget covering input, thinking and output; long decodes supported |
| Inputs / output | Text, image, audio and video in; text out. Knowledge cutoff March 2026 |
| Thinking levels | LOW, MEDIUM (default), HIGH |
| Free tier | None — "Gemini 4 Carbon is not available on the Free tier" (Google) |
API prices, per million tokens (Google, read November 13, 2026)
| Field | Detail |
| Standard | Input $2.00 and output $10.00 through November 30, 2026; $4.00 and $20.00 from December 1, 2026 |
| Batch / Flex | Input $1.00 and output $5.00 through November 30; $2.00 and $10.00 from December 1 |
| Priority | Input $3.60 and output $18.00 through November 30; $7.20 and $36.00 from December 1 |
| Cached input | $0.10 through November 30; $0.20 from December 1 (implicit caching only) |
| Note | Output prices include thinking tokens. The introductory rates apply to the Carbon model intended for public access (Google, launch post footnote) |
The Gemini 4 family
| Field | Detail |
| Gemini 4 Argon | Generally available October 13, 2026 — the efficient frontier; still Google's recommendation for most software and knowledge work; same introductory prices as Carbon |
| Gemini 4 Carbon | November 10, 2026 — the strongest model; long-horizon agentic work, complex coding, cybersecurity defense, scientific research |
| Gemini 4 Hydrogen | November 12, 2026 — the fast tier for everyone, all API tiers and consumer surfaces; flat $1.50 / $7.50 per million tokens across its 1M-token window |
Sources
- Google, official blog (Koray Kavukcuoglu, November 10, 2026): 'Gemini 4 Carbon is now available' — general availability, channels, introductory price and footnote, benchmark claims, /swarm in Antigravity, cybersecurity and science results, safety and honesty figures.
- Google, Gemini API pricing page (read November 13, 2026): Gemini 4 Carbon — not available on the Free tier; Standard $2.00 / $10.00 through November 30, 2026 and $4.00 / $20.00 from December 1, 2026; Batch, Flex, Priority and cached-input rates with the same change of date.
- Google Cloud documentation, Gemini 4 Carbon model page (updated November 11, 2026): model ID gemini-4-carbon, launch stage GA, release date November 10, 2026; unified 1,048,576-token limit; thinking levels; the cyber variant behind an allowlist on the Gemini Enterprise Agent Platform.
- Google, official blog (November 12, 2026): 'Introducing Gemini 4 Hydrogen' — the fast tier, its $1.50 / $7.50 flat pricing and its Artificial Analysis Intelligence Index figures against Argon.
- Artificial Analysis Intelligence Index, as reported November 12–13, 2026: Carbon (High) at 60 points against Claude Fable 5.2 (Max) at 65 — independent index readings, reported by third-party trackers of the index.
- Artwork: Google (official blog) — the Gemini 4 Carbon launch key art, and Google's lifecycle misalignment chart, both published with the November 10 announcement.