What Is Gemini 4 Carbon? Google’s Strongest Model, Explained

Gemini 4 Carbon is Google’s strongest AI model — the second model in the Gemini 4 series, after Argon — and as of November 10, 2026, it is generally available to developers and paying subscribers for the first time. Announced on October 28, Carbon spent its first two weeks restricted to a small group of cyber defenders. Now anyone with access to the paid Gemini API can build with it, and Google AI Ultra subscribers can use it in the Gemini app, in AI Mode in Search, and in Antigravity, Google’s agent-first development platform.

This guide explains what Carbon is, what it is designed to do, what it costs, how to access it, and how it compares with its closest rivals — Anthropic’s Claude Fable 5.2 and OpenAI’s GPT-6 Astra. All benchmark figures below are vendor-reported unless stated otherwise, and the comparison section explains where each model actually leads.

What is Gemini 4 Carbon?

Carbon is a frontier large language model from Google DeepMind, positioned above Gemini 4 Argon, which launched on October 13. Google describes it as its “strongest model with frontier performance in cybersecurity and long-horizon capabilities”. In plain terms, it is built for two things: hard engineering work that takes a long time, and finding and fixing security vulnerabilities in software.

The “long-horizon” part is the more important distinction. Most AI models are good at single, well-defined tasks — write this function, summarise this document. Carbon is designed for sustained agentic workflows: multi-step problems where the model has to keep going for hours or days, persist through dead ends, and finish a job rather than just start one. Google says its own teams have seen Carbon run for days to root-cause and patch subtle bugs across large codebases and low-level systems. That is the kind of work where a model that gives up, guesses, or quietly goes off-scope is worse than useless — and it is the capability Google is selling hardest.

Carbon is a proprietary, hosted model. Google distributes it through its own products and API. There are no downloadable weights, no open-weight licence, and no self-hosting option — a sharp contrast with open models such as Meta’s Muse Spark Open or Moonshot’s Kimi K3.1 Pro, which teams can run on their own infrastructure.

What Carbon is best at

Long-horizon coding and agentic work

Carbon’s headline strength is agentic software engineering: taking a complex task, breaking it into steps, using tools, checking its own work, and carrying the result through to completion. On Terminal-Bench 4.0, a test of terminal-based agentic coding, Carbon scores 65.2 percent at high effort — ahead of Claude Fable 5.2’s best listed run at 61.6 percent. On Zapier’s AutomationBench, which checks whether agents can complete assignments across real business applications, Carbon leads at 53.88 percent, ahead of Fable 5.2 at 51.45 percent.

It also improved sharply at building for the web. Google reports that Carbon jumped 89 elo points to fourth place on the WebDev Arena leaderboard, at 1768, with better output quality and better artifacts — the interactive pages and apps a model produces from a prompt — than Argon.

None of this makes Carbon the best model at every coding task. On FrontierCode 2.0, which grades changes to real repositories against the standard for merging them, Fable 5.2 leads at 49.1 percent against 40.6 percent for Carbon. The two models trade leads depending on the job, which is the honest picture: Carbon is strongest where the work is long, messy and tool-heavy, not on every benchmark row.

Cybersecurity: finding, exploiting and patching vulnerabilities

Cybersecurity is where Google claims its clearest lead, and where Carbon’s origin story matters. The model was first given to cyber defenders — not general users — through DeepMind’s Fairwind programme, and Google’s launch post is built around the defence lifecycle: find a vulnerability, prove it can be exploited, then fix it at the root cause.

On the three external benchmarks Google cites, Carbon sets state-of-the-art results. It scores 88.2 percent on SEC-bench Pro for autonomously finding real vulnerabilities buried in real-world software; 54.8 percent on ExploitGym for demonstrating that a vulnerability can actually be exploited; and 71.0 percent on CWE-bench for turning a validated finding into a working code fix. Google also reports that Carbon generates secure code at more than twice the success rate of Argon, and that its Chrome team found Carbon more than ten times better than another lab’s large frontier model at generating high-quality patches.

Two caveats belong here. First, Google ran those external cyber benchmarks without guardrails, so the scores describe the model’s raw capability, not what a normal user will see through the API. Second, the headline numbers are Google’s own, and no outside lab had reproduced them at the time of the general-availability announcement. What is independently documented is that the security platform Wiz used Carbon to find previously unknown vulnerabilities while scanning two major software providers in under two hours — though Google does not name the companies.

How it behaves: alignment and reliability

Google put unusual emphasis on how Carbon behaves, not just what it can do. The model builds on Argon’s safety stack — training-time safeguards, real-time misuse defences, prompt-injection resilience and chain-of-thought monitoring — and Google reports four measurable improvements over Argon:

• Fewer shortcuts. Carbon reduced successful cheating on Terminal-Bench 4.0 to 0.0 percent, with cheating attempts dropping to 0.4 percent, while improving overall accuracy. In live coding with Google’s own teams, high-severity misalignment alerts fell by almost 80 percent.
• More honest about uncertainty. In internal testing, Carbon cuts unsupported claims by about half compared with Argon. It puts methodological caveats directly into its deliverable, or withholds a blocked action with a clear list of what is still required, rather than fabricating or hedging around missing evidence.
• Better with ambiguity. Rather than guessing or forcing a path, Carbon retrieves the relevant sources and surfaces discrepancies to the user before acting.
• More reliable follow-through. Carbon reduced overreach by 26 percent and under-service by 29 percent — doing what was asked without doing extra things that were not asked, and without quietly dropping parts of the job.
• Stronger prompt-injection resistance. Where Argon led Gray Swan’s Indirect Prompt Injection Benchmark with a 0.7 percent attack success rate, Carbon scored 0.0 percent — resisting every evaluated attack attempt.

These are internal measurements, and they should be read that way. But they explain the design goal: a model that can be trusted to run for days inside a company’s systems without drifting off-task or inventing results is worth more than a model that is merely clever.

How much does Gemini 4 Carbon cost?

Carbon launched at the same introductory price as Argon: $2 per million input tokens and $10 per million output tokens. That introductory rate runs through November 30, 2026, after which the price doubles to $4 per million input tokens and $20 per million output tokens. Cached input tokens — repeated prompts such as the same codebase or policy documents — are priced at 95 percent off the input rate, which works out at about $0.10 per million cached input tokens during the promotion and $0.20 afterwards.

For the full, current price list, see Google’s Gemini API pricing page.

The price is Carbon’s sharpest competitive weapon. At the introductory rate, it costs a quarter of Claude Fable 5.2’s $8 per million input and $40 per million output tokens, and a fifth of OpenAI’s GPT-6 Astra at $10 and $50. Even after the November 30 increase, Carbon at $4 and $20 is half the price of Fable 5.2 and equal to GPT-6 Astra’s standard rate — while, on several agentic benchmarks, it scores close to or ahead of both.

But the token price is not the job price. Artificial Analysis estimates an average of $7.76 per task for Carbon at high reasoning effort on its index suite, against $26.54 for Fable 5.2 at maximum effort — yet Fable 5.2 at high effort costs $9.08 and scores two points higher than Carbon. And on Zapier’s AutomationBench, once failed attempts are counted, Carbon’s cost per passing task is $3.11 at medium effort, against $3.94 for its high-effort setting. The practical lesson is to measure cost per finished job, including retries and human review time, rather than comparing token prices alone.

How to access Gemini 4 Carbon

Since November 10, 2026, there are three ways in:

• Developers: the paid tier of the Gemini API, through Google AI Studio and Android Studio. The announcement points developers to the Gemini API developer guide for model details and setup.
• Businesses: Gemini for business, with API and model-picker access, plus specialised agents such as CodeMender and AlphaEvolve.
• Consumers: Google AI Ultra subscribers — $99.99 a month for Ultra 5x and $199.99 a month for Ultra 20x in the US — in the Gemini app, in AI Mode in Search, and in Antigravity.

Google has not published a country-by-country rollout schedule, and it has not confirmed Vertex AI availability, so enterprises that buy through Google Cloud contracts should check before planning a migration. Google is also adding Carbon to a programme that gives selected scientists six months of free access to Antigravity’s highest usage tier.

In Antigravity, Google is previewing a feature called /swarm: a single prompt becomes a coordinated team of agents that divide up a project and check one another’s work, with Carbon assigned to architecture, coordination or audit work while lighter models handle parallel implementation tasks in isolated workspaces. The idea is the same one Anthropic describes for Fable 5.2 — use the strongest model as the orchestrator and cheaper models for the routine steps. For background on how these systems work, see our guide to agentic AI.

Gemini 4 Carbon vs Claude Fable 5.2 vs GPT-6 Astra

The three frontier models released within weeks of each other now define the top of the market, and the choice between them depends on the job rather than a single leaderboard. For the full picture on Anthropic’s model, see our guide to Claude Fable 5.2.

On Artificial Analysis’s Intelligence Index v5.0, Carbon debuted in first place at 59 when it was announced on October 28. By the time of its general availability, Fable 5.2 had taken the lead at 65 points at maximum effort, with Carbon at 60 at high effort, Claude Opus 5.5 at 59 and GPT-6 Astra at 57. The gap is real but narrow, and it changes with the task.

Carbon leads on Terminal-Bench 4.0 (65.2 percent against Fable’s 61.6 percent) and on AutomationBench (53.88 percent against 51.45 percent), and it leads on the cyber benchmarks by a wide margin. Fable 5.2 leads on FrontierCode 2.0 (49.1 percent against 40.6 percent), on Terminal-Bench-Science (74.6 percent against 64.8 percent for Carbon), and on the WANDR research benchmark. GPT-6 Astra sits close behind both on most rows while costing two and a half times Carbon’s introductory token price — though its cost per completed task can be lower than either rival on some workloads, because it needs fewer tokens and retries to finish.

On price, Carbon wins outright at the token level: $2 and $10 through November 30, then $4 and $20, against Fable 5.2 at $8 and $40 and Astra at $10 and $50. For a team running long agentic jobs, that difference compounds quickly — which is exactly the market Google is aiming at.

There is one structural difference worth noting. Anthropic splits its model into two doors: Fable 5.2 for general use and Mythos 5.2 for verified organisations, with stronger cyber and biology safeguards on the general version. Google does the same thing in a different shape: the general-use Carbon available since November 10, and a separate Carbon for cyber that remains gated behind the Fairwind programme for vetted defenders. A security team comparing the two has to start with which version it can actually access.

Carbon vs Argon: what changed?

Carbon is the second model in the Gemini 4 series, announced four weeks after Argon and two weeks after Argon reached the public. Compared with Argon, Google claims significant gains across coding and agentic tasks, a large jump in web development quality, and the cybersecurity lead described above — plus the alignment improvements: fewer shortcuts, fewer unsupported claims, better handling of ambiguity and better follow-through.

For most software engineering and knowledge work, Google itself recommends Argon, which remains available at the same $2 and $10 introductory price. Carbon is the model for the hardest jobs: long-horizon engineering, security-sensitive work, and tasks where a model needs to keep going for days without losing the thread. If your work does not need that, Argon is cheaper to run at the same token price only because it uses fewer tokens on simpler work — the two models cost the same per token, so the choice is about capability, not price.

Carbon also differs from Google’s lighter models in the family. Gemini Omni 1.2 Flash, launched the same week, is a fast multimodal model built for short video and everyday tasks — the opposite end of the range from Carbon’s days-long engineering runs.

What Google has not said

A careful reader of the launch materials will notice the gaps. Google has published no Carbon model card, no context-window or output-limit specification — Artificial Analysis lists a 1 million-token context window — no speed figure, and no Carbon-specific safety report. The safety report linked from the announcement is Argon’s 217-page report, which does not mention Carbon. The only Carbon-specific document Google has published is a five-page evaluation methodology note with the benchmark table.

Google also has not said when Carbon will reach free-tier users, if ever, or when it will be available in every country. The staggered rollout — defenders first, then developers and Ultra subscribers — is deliberate: Google says it is gathering feedback and iterating on guardrails before wider release, and it took part in the US government’s voluntary pre-release model testing process before launch.

Who should use Gemini 4 Carbon?

Carbon makes most sense for three groups. Software teams running long agentic jobs — large refactors, bug hunts across big codebases, multi-day engineering tasks — where persistence and reliable follow-through matter more than raw speed. Security teams that need to find, validate and patch vulnerabilities, and that can access the model through the API or the Fairwind programme. And cost-conscious teams that want frontier-level agentic performance at a fraction of Fable 5.2’s or Astra’s token price, and are willing to measure cost per finished job rather than trust the leaderboard.

It makes less sense for everyday chat, short writing tasks, or quick code completions, where lighter models such as Gemini 4 Argon or GPT-6 Cyber‘s siblings are faster and cheaper per job, and for teams that need to run inference on their own infrastructure, where open-weight models are the only option.

Frequently asked questions

Is Gemini 4 Carbon free?
No. Carbon is available through the paid Gemini API and to Google AI Ultra subscribers. The introductory price is $2 per million input tokens and $10 per million output tokens through November 30, 2026, doubling to $4 and $20 afterwards. There is no free tier announced.

When did Gemini 4 Carbon become available?
Google announced the model on October 28, 2026, with access limited to selected cyber defenders. It became generally available on November 10, 2026, to developers through the paid Gemini API and to Google AI Ultra subscribers.

How is Carbon different from Argon?
Carbon is the stronger, second model in the Gemini 4 series. It improves on Argon in long-horizon agentic work, web development and cybersecurity, and it is more reliable in follow-through and honesty about uncertainty. Both models share the same introductory token price; Google recommends Argon for most everyday engineering work.

Is Carbon better than Claude Fable 5.2?
It depends on the task. Carbon leads on Terminal-Bench 4.0 and AutomationBench and on every cyber benchmark Google cites, and it costs a quarter of Fable 5.2’s token price through November 30. Fable 5.2 leads on the broad Artificial Analysis index and on coding and science benchmarks such as FrontierCode 2.0. For long agentic and security work, Carbon is the stronger value; for general coding and research, Fable 5.2 currently leads.

Can I run Carbon on my own servers?
No. Carbon is proprietary and hosted only, distributed through Google’s API and products. There are no downloadable weights.