Two weeks ago, Google showed the world Gemini 4 Carbon but would not let most people touch it. On November 10, 2026, that changed. Google DeepMind opened Carbon — the model it calls its strongest yet — to paying developers, enterprises, scientists and consumers, ending a two-week preview that had been limited to cybersecurity defenders in its Fairwind Program. Carbon is built for the jobs other models still fumble: multi-hour software engineering, vulnerability hunting, and research that spans dense documents, charts and code. Here is what it is, what it costs, and when it is actually worth using over its sibling, Gemini 4 Argon.
What Gemini 4 Carbon is — and who it is for
Google positions Carbon as the heavyweight of the Gemini 4 family. Where Argon is the everyday frontier model, Carbon is the one you reach for when a task runs for hours rather than minutes: long-horizon agentic workflows across complex knowledge work, software engineering and scientific reasoning, plus frontier cybersecurity capabilities. That is the framing from Koray Kavukcuoglu, SVP at Google DeepMind and Google’s Chief AI Architect, who announced the release on November 10.
The model had a staged rollout. Google announced Carbon on October 28, 2026, but for the first 13 days only vetted cybersecurity defenders could use it through the Fairwind Program, where Google worked with impacted organisations on live remediation. The November 10 release opens a general-use build to a much wider audience: paying developers and enterprises through the Gemini API in Google AI Studio and Android Studio, and through Gemini for business. Google AI Ultra subscribers also get it in Google Antigravity, the Gemini app, and AI Mode in Search. The model ID in the API is gemini-4-carbon.
There is a catch worth stating plainly. Carbon is not on the free API tier, and in the Gemini app it is Ultra-only for now. Google has given no date for free or lower-tier access, so for most individual users the model remains behind a paywall — a paid API key or a $99.99-a-month Ultra subscription.
How much it costs — and the price rise coming on December 1
Carbon launches at the same introductory price as Argon: $2 per million input tokens and $10 per million output tokens, with cached input at $0.10 per million tokens. But that rate has a firm expiry date. After the introductory period ends on November 30, 2026, the price doubles to $4 per million input tokens and $20 per million output tokens, with cached input rising to $0.20.
That deadline is part of the decision, not a footnote. A team that builds a workflow on Carbon at the introductory rate should budget for the standard price from December 1, and Google’s own framing of the model as a premium, token-hungry option for the hardest jobs makes the arithmetic matter: Carbon is designed for tasks where the model thinks long and calls tools repeatedly, so token consumption per task is high even when the per-token price looks modest.
The benchmark claims — impressive, but Google’s own numbers
Google says Carbon ranks first on three independent workflow benchmarks — CursorBench 4.0, Perplexity’s WANDR and Zapier’s AutomationBench — at more than twice lower cost per task than the next best external model. The most concrete figure is AutomationBench: on 657 tasks, Carbon scored 53.9% at High effort, ahead of Argon at 51.3%, Claude Sonnet 5.5 at 44.7% and GPT-6 Astra at 41.4%. On CursorBench 4.0, the coding-agent benchmark, Carbon leads with just under 60%, ahead of Claude Fable 5.2 and Argon. On the Artificial Analysis Intelligence Index v5, Carbon scores 59, ahead of Claude Opus 5.5 at 58 and GPT-6 Astra at 57, and on Terminal-Bench 4.0 it reaches 65.2%.
Google also launched a new benchmark with the release: the Multimodal Knowledge Agent (MKA) test, built with Artificial Analysis, which measures agents reasoning over dense files with charts, images, blueprints, schematics and tables. Carbon finishes second there at 54%, just behind Claude Fable 5.2 at 57%, but at what Google says is more than six times lower average cost per task than the leader.
Every one of those numbers comes from Google or from benchmarks Google helped build, and independent replication had not landed at publication time. The AutomationBench lead over Argon is also narrow — roughly 17 tasks out of 657 — which may sit within the noise of a single run. Read the rankings as a signal about where Google thinks Carbon wins, not as settled proof. The cost-per-task framing is the more interesting claim: Google is arguing Carbon wins on value, not just on raw scores, and that is a claim independent testers can check quickly once they have access.
Cybersecurity is the model’s real showcase
The strongest case for Carbon is not general coding — it is security work. Google says Carbon leads on SEC-bench Pro at 88.2% for finding vulnerabilities, ExploitGym at 54.8% for checking whether they are exploitable, and CWE-bench at 71.0% for patching them. On the Gray Swan prompt-injection benchmark, Carbon showed a 0.0% attack success rate, down from 0.7% for Argon — though no model is fully resistant at scale, and Google is careful to note that the security figures reflect the version verified defenders reach through Fairwind.
The same day as the release, DeepMind showed what that capability looks like in practice. Its new AlphaProof Code framework, powered by Gemini, formally verified key security properties of the actual code behind Signal’s cryptographic protocol — generating, by one report, more than six million lines of machine-checked proof over four weeks, and correctly rejecting older versions with known bugs. That is a different kind of evidence from a benchmark table: a real codebase, a real protocol used by millions of people, and proofs a machine can check rather than a score a vendor reports.
For security teams, the practical implication is narrower than the headline. Carbon is a tool for finding, validating and patching vulnerabilities in code you already own, and its Fairwind heritage means it was tested on live incidents before the public release. It does not replace a security review, and its own makers warn that a model this capable at finding flaws is also a model that could help attackers — which is exactly why the cyber-defender version stays separate from the public build.
/swarm: Carbon arrives with a team of agents attached
Carbon does not launch alone. Google is previewing /swarm inside Antigravity, its agent-first development platform, for Ultra subscribers. The feature turns one prompt into a coordinated team of AI agents that split a complex task in real time — one agent planning, others writing code, testing and reviewing in parallel. It rolls out in preview across the Antigravity app and CLI, with Carbon as the model powering it.
This is the part of the launch that most directly changes how developers work. A single powerful model is a better pair programmer; a swarm of them is a different workflow, closer to managing a small team than prompting a chatbot. The open question is coordination overhead: multi-agent systems can spend as much effort talking to each other as doing the work, and /swarm is still a preview. But paired with Carbon’s long-horizon design, it is Google’s clearest bet yet that the next step in coding tools is not a smarter single agent, but a managed team of them.
Carbon vs Argon: which Gemini 4 model should you use?
Google is explicit that Carbon complements Argon rather than replacing it, and the comparison table makes the split clear. Argon is the model for most work — faster, cheaper per token, and good enough for everyday coding and writing. Carbon is for the hardest multi-step jobs: the migration that touches a hundred files, the vulnerability hunt across a whole codebase, the research task that needs the model to keep going for hours without losing the thread.
The price makes that choice for you in most cases. Until November 30, both models cost the same per token, so there is little reason not to try Carbon on a hard task. From December 1, Carbon costs twice as much per token as the introductory rate, and its long-horizon style means it will also consume more tokens per task. Use Argon as the default, Carbon as the specialist — and measure the cost per completed task, not the cost per token, before committing a workflow to either.
That pattern is now familiar across the industry. Google’s own Gemini family is splitting the same way its rivals’ lineups are: a fast, cheap model for the everyday, and a heavy, expensive one for the hardest work. The same week brought Mistral Large 4, a trillion-parameter open-weight model aimed at teams that want to own their frontier model outright, and NVIDIA Cosmos 3 Nano, open world models that let robots rehearse in simulation — different answers to the same question of who controls the most capable AI, and at what price.
What to watch next
Three things will tell us whether Carbon lives up to the launch. First, independent benchmarks: once researchers outside Google run Carbon on the same tasks, we will know whether the AutomationBench and CursorBench leads survive outside Google’s own reporting. Second, the December 1 price test: when the introductory rate doubles, developers will find out whether the swarm is worth the tokens. Third, the Flash variant: reports suggest a lighter Gemini 4 Flash is likely to follow, which would complete the family and give Google an answer at every price point.
For now, the practical steps are simple. Paying developers can start building in Google AI Studio with the model ID gemini-4-carbon; Ultra subscribers can try it in Antigravity, the Gemini app and AI Mode in Search. Try it on a genuinely hard, multi-step task — a codebase migration, a security audit, a research synthesis — and track what it costs per finished job. That is the test Google is inviting, and it is the only one that matters.
Bottom line
Gemini 4 Carbon is Google’s strongest model, and for the first time anyone with a paid API key or an Ultra subscription can use it. The benchmarks are impressive but vendor-reported, the cybersecurity story is the most concrete part of the launch, and the price doubles on December 1 — so the next three weeks are the cheap window to find out whether Carbon earns its premium. The model cycle is spinning faster than ever, and Carbon is Google’s answer to a simple question: what happens when the agent does not stop after one turn, but keeps working for hours? Now developers can find out for themselves.
Related reading and sources
- Google’s official Gemini 4 Carbon announcement
- Google’s developer guide for building with Carbon
- Google Antigravity on Carbon and the /swarm preview
- Our coverage of Google’s Gemini Omni 1.2 Flash video model
- Mistral Large 4, the trillion-parameter open-weight rival
- OpenAI’s acquisition of Pierre, Git infrastructure for AI-written code
- NVIDIA Cosmos 3 Nano, open world models for robot simulation
