 |
| AI-generated illustration of a glowing neural-network core with smaller agent orbs working on holographic code windows and charts. |
On 10 November 2026, Google made Gemini 4 Carbon — the model it calls its strongest yet — available beyond the small group of cybersecurity defenders who had exclusive access to it for two weeks, opening it to paying developers, enterprises, and Google AI Ultra subscribers. It arrives with a striking set of claims: first place on three workflow benchmarks, a new /swarm feature that turns a single prompt into a team of coding agents, and cybersecurity results that Google says compressed weeks of vulnerability research into a single afternoon. It also arrives with limits that matter as much as the claims. Most people cannot use it at all, the introductory price doubles on 1 December, and almost every number in the announcement comes from Google's own testing rather than independent verification. This article explains what actually changed on 10 November, what Carbon is and is not, and whether it is worth your money or your attention.
Gemini 4 Carbon is the second model in the Gemini 4 family, following Gemini 4 Argon, which reached paying users in mid-October 2026. Google describes Carbon as built for the tasks that test the limits of a model's intelligence: complex, multi-step coding and debugging, open-ended scientific research and discovery, long-horizon agentic work, and cybersecurity. It deliberately uses more tokens than Argon to explore a problem thoroughly, reason deeply, and run extended verification loops. Argon, Google says, remains the ideal choice for the majority of software engineering and knowledge work, while Carbon is for the hardest jobs. Carbon was first announced on 28 October, but only for a select group of cybersecurity defenders through Google's Fairwind Program. The 10 November news is therefore not a new model — it is the gate opening. The model card gives Carbon a 1 million token context window and a March 2026 knowledge cutoff, and there is no free tier: free API users and AI Pro subscribers do not get Carbon, and Google has removed Gemini 3.1 Pro from the Gemini app as part of the same reshuffle.
As an Amazon Associate, we earn from qualifying purchases.
Prices and savings can change. Check the retailer for the current price.
What actually changed on 10 November
The change is about who can reach the model, and where. From 10 November, Carbon is available through the paid tier of the Gemini API in Google AI Studio and Android Studio, and through Gemini for business for enterprises. Google AI Ultra subscribers — the company's top paid plan, starting at $99.99 a month in the United States — can use it in the Gemini app, in AI Mode in Search, and in Google Antigravity, the company's agent-based coding tool. Google's chief executive, Sundar Pichai, announced the wider release on X the same day, and the launch was covered the same day by 9to5Google, The Verge, and a string of other outlets. Scientists get a separate route: Google opened applications four weeks ago to give 10,000 scientists free access to Argon in Antigravity for six months, says it has already onboarded thousands of them, and is now expanding that programme to Carbon as well, with applications still open and onboarding on a weekly rolling basis. For cybersecurity defenders, Carbon for cyber remains available by application through the Fairwind Program. What has not changed is the price of entry for everyone else. If you are on the free tier or on AI Pro, nothing about this announcement gives you Carbon, and Google has not said when — or whether — that will change.
The benchmarks, and how much to trust them
Google's headline numbers are strong, and they deserve a careful reading. Carbon ranks first on CursorBench 4.0, Perplexity's WANDR, and Zapier's AutomationBench — three benchmarks designed to test models on real workflows rather than exam-style questions — at what Google describes as more than twice the lower cost per task than the next best external model. On a new Multimodal Knowledge Agent benchmark that Google built with Artificial Analysis to test agentic work over dense files with charts, images, blueprints, and tables, Carbon ranks second while costing more than six times less per task on average than the top-performing model. In pre-release evaluations it matched the best model tested on Hebbia's end-to-end financial research evaluation, and ranked at the top out of 23 frontier model configurations on Balyasny Asset Management's General Financial Intelligence benchmark, which spans legal expertise, agentic tool use, coding and data engineering, and perception. These are genuinely good results, and the financial benchmarks in particular are the kind of independent, domain-specific tests that are harder to game than a general leaderboard. But two cautions apply. First, the benchmarks are ones Google chose to highlight, and "cost per task" depends heavily on how many tokens a model burns to get there — Carbon is designed to use more tokens than Argon, so a lower price per token does not automatically mean a lower bill. Second, on the best-known independent index, the Artificial Analysis Intelligence Index, Carbon is not the leader. Anthropic's Claude Fable 5.2, released on 9 November — the day before Carbon's wider release — holds the top score at 65, while Carbon sits around 59 to 60 on Google's own benchmark table. The honest picture is a close race in which the two models lead on different measures, not a single runaway winner, and the next round of independent testing will say more than either company's launch post.
/swarm: one prompt becomes a team of agents
The most interesting part of the announcement may not be the model itself but what Google is previewing alongside it. Called /swarm, it lives in Google Antigravity and turns a single prompt into a coordinated team of AI agents that you can watch in real time. Typing /swarm in the composer hands the prompt to a coordinator agent, which spins up parallel worker agents in separate copies of the code so their edits never collide, plus adversarial verifier agents that run tests and challenge every change. Developers can assign Carbon to the roles that demand the deepest reasoning — architectural design, coordination, and auditing — while pairing it with faster, lighter models for routine work across parallel isolated workspaces. Google says its own engineers used /swarm for five weeks of internal testing on multi-package migrations, cross-subsystem root-cause investigations, and complex architectural redesigns, and its demo shows a swarm building a 3D game in which a pelican rides a bike around Google's Mountain View campus. The feature is rolling out in preview in the Antigravity app and command-line tool, and it is still a preview: Google has not explained in detail how swarms share context between agents, handle conflicting changes, or decide which tasks need Carbon instead of a lighter model. There is also a warning worth taking seriously, and it comes from Google itself: swarms consume quota much faster than a standard chat. A team of agents doing parallel work burns tokens in parallel, so anyone trying /swarm on a large project should set limits first, because the bill can grow far faster than a single chat ever would.
The cybersecurity case is the strongest claim — and the one to watch
If there is one area where Carbon's launch is more than a benchmark press release, it is cybersecurity, and that is also where the stakes are highest. Carbon for cyber is the version of the model that was available to defenders first, and Google's examples are concrete. In Project Fortify, Google's autonomous system spanning scanning, triage, and remediation, Carbon hit 100 percent recall on scanning and 95 percent precision on triage on Google's internal benchmark, and on open-source repository testing it showed over 25 percent relative improvement on Argon's already strong recall rate, surfacing previously undetected, real-world logic flaws. Google's Cloud Vulnerability Research team used Carbon to prove that two bugs in a popular open-source software package were in fact exploitable vulnerabilities, comparable in severity to the recent PHP2Shell exploit, helping maintainers re-classify the bugs as critical and issue a patch. As part of Wiz's Scan for Good programme, a team used Carbon to assess vulnerabilities in public digital infrastructure and uncovered zero-day vulnerabilities in two major software-as-a-service providers in under two hours, compressing weeks of vulnerability research into a single afternoon, then worked with the affected organisations to remediate them and mitigate potential impact on tens of thousands of companies. The safety side of the same coin matters too. Because Carbon is a stronger cyber model, Google says it strengthened its safeguards in proportion. In simulated cyberattack testing environments, Carbon blocked 99 percent of attack trajectories from an adaptive red-teaming agent, and the per-attempt breach rate fell to 0.8 percent, down from 3.1 percent on Argon. Automated cyberattack query success fell to 2.17 percent in Gray Swan's testing, compared to 9.86 percent for Argon, and Carbon achieved a 0 percent prompt injection attack success rate on Gray Swan's indirect prompt injection benchmark — the first time that milestone has been reached, according to Google. External testing by Trajectory Labs found zero new universal jailbreaks across 132 hours and over 13,000 requests. These are impressive figures, but they are still Google's tests, run in seeded simulated environments, and the offensive-use risk is exactly why access to Carbon for cyber is gated behind an application process. Independent replication, and the question of how these defences hold up against attackers who adapt, will decide how much of this survives contact with the real world.
Science: the Signal proof and the laboratory bench
The scientific claims are the part of the announcement that reaches furthest beyond coding. Google DeepMind researchers built a new agentic framework for formal verification called AlphaProof Code and used it, with Carbon alongside other models, on a major security challenge: mathematically proving the actual code underlying the cryptographic protocol of Signal, the messaging app used by hundreds of millions of people. The researchers translated security goals into mathematical theorems, and AlphaProof Code agents then used Carbon to generate millions of lines of code in Lean, a computer language for writing mathematical proofs, helping them prove key security properties in weeks — a task that historically takes expert teams years. The result points to scalable formal verification, long considered a gold-standard approach to building secure software, becoming practical rather than theoretical. In the laboratory, Carbon went beyond standard analysis for drug toxicity research company Axiom Bio, finding novel mechanisms of toxicity the company had not found before; Axiom is now moving its production scientific analyses and agents to both Carbon and Argon. Given only lab notebooks and raw data, Carbon surfaced hard-to-interpret data anomalies, diagnosed subtle wet-lab experimental artefacts such as incomplete automated mixing in deep wells, improved methods to deconvolute chromatography peaks, and caught trace sample carryover in research instruments. In protein structure prediction, where models like AlphaFold 3 struggle on understudied proteins with few evolutionary relatives, researchers used Carbon in a multi-agent setup to autonomously discover remote homologs across petabase-scale unannotated sequence archives, increasing the number of homologs retrieved by 40-fold on average across seven low-homolog protein families from the CASP17 blind test set. Google is careful to add that broader benchmarking is still underway there, which is the right caveat: a 40-fold improvement in retrieval is a promising upstream fix for structure prediction, not yet a proven improvement in the final structures themselves.
What it costs — and the catch
The pricing is simple to state and easy to underestimate. Carbon costs $2 per million input tokens and $10 per million output tokens until 30 November 2026, and from 1 December the price doubles to $4 and $20. That matches the introductory pricing Gemini 4 Argon carried at launch, and Argon's price roughly doubled after its own introductory period, so the pattern is established. Three things make the real cost higher than the headline rate. First, reasoning tokens count on every bill, and Carbon is designed to think longer than Argon. Second, /swarm multiplies usage by running several agents in parallel, and Google warns that it burns quota much faster than a normal chat. Third, the introductory rate is explicitly temporary, so anyone building a workflow on Carbon should budget at the December price, not the November one. For everyday Gemini users the price is, for now, irrelevant: Carbon is not available on the free tier, and AI Pro subscribers keep Argon. The only no-extra-cost route is being an AI Ultra subscriber, at $99.99 a month in the United States, where Carbon is included in the Gemini app, AI Mode in Search, and Antigravity. Scientists can apply for the free six-month Antigravity tier, and cybersecurity defenders can apply to the Fairwind Program. Everyone else pays per token, and should test Carbon against Argon on their own real tasks before 1 December, when the price of that test doubles.
Checklist: should you try Gemini 4 Carbon?
- If you are a Google AI Ultra subscriber, try it now in the Gemini app or Antigravity — it costs you nothing extra, and it is the fastest way to learn whether the deeper reasoning helps your work.
- If you are a paying API developer with long-horizon coding, security, or research work, run Carbon and Argon side by side on your own tasks and compare the cost per finished task, not the price per token.
- If you are on the free tier or AI Pro, nothing changes for you — Argon remains your model, and there is no announced date for Carbon to reach you.
- If you are budgeting, plan at the 1 December price of $4 input and $20 output per million tokens, and remember that Carbon uses more tokens than Argon by design.
- If you want to try /swarm, set usage limits first — parallel agents burn quota far faster than a single chat.
- If you do security research, apply to the Fairwind Program for Carbon for cyber rather than waiting for general access.
- If you are a scientist, apply for the free six-month Antigravity tier, which now covers both Argon and Carbon.
- If you need independent proof before switching, wait for third-party benchmarks — the launch numbers are Google's own, and the Fable 5.2 comparison shows the race is still open.
Bottom line
The 10 November announcement was a real expansion, not a new model: Gemini 4 Carbon has existed since 28 October behind a gate for cybersecurity defenders, and the gate has now opened for paying users. Its strongest claims are in cybersecurity and science, where Google has concrete examples — zero-days found in hours, the Signal protocol formally verified in weeks, novel toxicity mechanisms found for a drug research company — rather than in the benchmark tables, where the leadership is real on Google's chosen measures but not on the best-known independent index, where Claude Fable 5.2 currently leads. For AI Ultra subscribers, trying Carbon costs nothing extra and is the obvious move. For API users, the sensible test is a side-by-side run against Argon on real work, budgeted at the December price. For everyone else, this is a launch to watch: the model is genuinely capable, the /swarm preview points to where coding assistants are heading, and the next few weeks of independent testing will show how much of Google's announcement holds up.
Reference sources
- Google, "Gemini 4 Carbon is now available", 10 November 2026
- Google Antigravity, "Gemini 4 Carbon and /swarm", 10 November 2026
- 9to5Google, "Gemini 4 Carbon rolling out to Gemini app for AI Ultra, Argon remains", 10 November 2026
- ShiftDelete, "Gemini 4 Carbon Pricing: $2 Now, $4 From Dec. 1", 10 November 2026
- VectorWire, "Google Opens Gemini 4 Carbon to Paid Users, Previews Swarm", 10 November 2026
- MadRobot, "Gemini 4 Carbon is out: price, /swarm and who can use it", 10 November 2026
- TechResearchOnline, "Gemini 4 Carbon Opens to Developers and Enterprise Teams", 10 November 2026
Leave a public comment