Google has a new number one artificial intelligence model, and almost nobody can use it yet. On October 28, 2026, Google DeepMind announced Gemini 4 Carbon, calling it its strongest model so far, built for cybersecurity defense and long, multi-step engineering work. Independent evaluator Artificial Analysis immediately placed it first on its Intelligence Index v5.0 with a score of 59, ahead of Anthropic's Claude Opus 5.5 at 58 and OpenAI's GPT-6 Astra at 57.
That headline comes with a catch that defines the whole story: Carbon is not on sale, not in the Gemini app, and not on the public API. Google is handing it first to a small group of vetted cyber defenders through its Fairwind Program, and everyone else has no date for access at all. So this article answers the questions people are actually asking: what Carbon is, how it compares with the models before it and beside it, who it is for, what it costs, where it has launched, and what you should use instead while you wait.
What Gemini 4 Carbon actually is
Carbon is the second model in Google's Gemini 4 family. It follows Gemini 4 Argon, which Google first limited to cybersecurity partners on September 30, 2026 and then made generally available on October 13. Two weeks later came Carbon, built on Argon's foundation and tuned for a specific job: the full cybersecurity defense lifecycle — finding vulnerabilities, proving they are real by exploiting them, and then writing the patch — plus engineering tasks that run for hours or days rather than minutes.
Google DeepMind chief Koray Kavukcuoglu described Carbon as "our strongest model with state-of-the-art cybersecurity and long horizon capabilities." The phrase "long horizon" is doing real work here. Most AI coding assistants are good for a short burst and then lose the thread as context piles up: they cut corners, invent API parameters, or declare a task impossible. Google says its internal teams watched Carbon persist through dead ends for days while tracing subtle bugs across large codebases and low-level systems. The company published no completion rate for those tasks, so treat that as a claim, not a measurement.
The most vivid independent evidence for long-horizon stamina is Vending-Bench 2, a test that gives a model $500 in seed capital and asks it to run a simulated vending-machine business for 365 days: negotiating with suppliers, restocking perishables, renegotiating rental contracts, and adapting to shifting demand. Previous frontier models tended to drift badly around day 90, draining cash through over-ordering or failing to secure bulk pricing. Carbon finished with a bank balance of $16,484, ahead of GPT-6 Astra at $15,515 and well ahead of Claude Opus 5.5 at $9,235. A simulated vending business is not a real company, but as a stamina test it is hard to fake: the model has to keep making coherent decisions across a year of simulated state.
On the technical side, independent evaluator listings give Carbon a 1 million-token context window, with text, image, video, and speech inputs and text output. Those figures come from Artificial Analysis and Vals listings; no Carbon-specific Google developer specification had appeared at announcement time. Google has not disclosed Carbon's parameter count or cutoff, and there are no downloadable weights — Google-hosted only.
The headline numbers
Start with the ranking that made the news. Artificial Analysis rebuilt its Intelligence Index as v5.0 with harder private test sets and agentic environments, and Carbon took the top spot:
The jump over Argon is the striking part: 50 to 59 in a single generation step, two weeks after Argon's general release. Artificial Analysis also noted where the gains came from, with Presentation up 293 Elo points to 1598 and Analytical Quality up 120 points to 1671. On its Coding Agent Index, Carbon scored 59, three points behind Claude Code with Opus 5.5 — a reminder that the overall crown and the coding-agent crown are different contests.
Then there is the cybersecurity trio, which is the real reason Carbon exists. All three scores below are Google-reported, and the security evaluations were run without guardrails, which matters for how you read them:
| Security benchmark | What it measures | Carbon | Best rival |
|---|---|---|---|
| SEC-bench Pro | Finding vulnerabilities | 88.2% | GPT-6 Astra 85.4% |
| ExploitGym | Proving flaws by exploiting them | 54.8% | GPT-6 Astra 42.4% |
| CWE-bench | Writing the patch | 71.0% | Argon and Astra, both 68.0% |
Google's internal claims go further: Carbon generates secure code at more than twice Argon's success rate, cut high-severity misalignment alerts in live internal coding by almost 80%, roughly halved unsupported claims compared with Argon, and reduced successful cheating on Terminal-Bench 4.0 to zero, with cheating attempts at 0.4%. On Gray Swan's Indirect Prompt Injection Benchmark, attack success fell from 0.7% for Argon to 0.0% for Carbon. And the Chrome team's internal test reportedly found Carbon producing high-quality patches more than 10 times better than another lab's large frontier model — a claim Google published without naming the rival, so it cannot be checked. None of these internal figures have been reproduced outside Google yet.
One more number deserves attention because it shows the scale of the long-context improvement. On GraphWalks, a test that runs from 256,000 to 1 million tokens of context, Carbon scored 96.2%, against 84.2% for Argon and 71.8% for GPT-6 Astra. That is a 12-point jump over Argon and roughly a 24-point gap over GPT-6 Astra.
Carbon vs Argon vs GPT-6 Astra vs Claude Opus 5.5: the full comparison
Google published a 17-row comparison in its announcement, with Carbon run at its highest reasoning setting and most rival numbers self-reported by each provider. Here it is in full, with the leader in each row marked. Two rows break Google's way: FrontierSWE 2, where GPT-6 Astra leads, and OSWorld-2.1, where Claude Opus 5.5 leads.
| Benchmark | Carbon | Argon | GPT-6 Astra | Claude Opus 5.5 | Claude Fable 5.1 |
|---|---|---|---|---|---|
| AA Intelligence Index v5.0 | 59 | 50 | 57 | 58 | 49 |
| Vals Index v2.2 | 70.4% | 69.2% | 65.0% | 70.2% | — |
| SWE-sweep (agentic coding) | 9.6% | 4.3% | 9.5% | 3.6% | 2.9% |
| FrontierSWE 2 | 61.2% | 55.0% | 65.5% | 62.3% | 56.3% |
| Terminal-Bench 4.0 | 65.2% | 57.1% | 59.1% | 59.6% | 52.0% |
| Agents' Last Exam v1.1 | 77.4% | 69.0% | 73.1% | 71.6% | — |
| OSWorld-2.1 (computer use) | 81.5% | 74.7% | 80.9% | 87.2% | 76.8% |
| AutomationBench (knowledge work) | 53.9% | 51.3% | 41.4% | 42.5% | 31.4% |
| Vending-Bench 2 (final balance) | $16,484 | $13,718 | $15,515 | $9,235 | $5,422 |
| Chartography (multimodal) | 72.5% | 71.6% | 71.0% | 66.3% | 46.2% |
| Blueprint-Bench 2 (multimodal) | 58.6% | 54.4% | 49.7% | 51.2% | 41.9% |
| GraphWalks (256k–1M context) | 96.2% | 84.2% | 71.8% | 66.8% | 65.0% |
| RiemannBench (science and math) | 80.0% | 76.0% | 72.0% | 69.6% | 65.6% |
| Terminal-Bench-Science 0.1 | 64.8% | 52.9% | 63.3% | 59.0% | 43.3% |
| CWE-bench (patching) | 71.0% | 68.0% | 68.0% | 67.0% | 58.0% |
| ExploitGym (exploitation) | 54.8% | 42.0% | 42.4% | 34.5% | — |
| SEC-bench Pro (finding bugs) | 88.2% | 83.4% | 85.4% | — | — |
Read the table as a map of where each lab is strongest. Carbon leads on breadth: coding, knowledge work, multimodal understanding, long context, science, and all three security tests. OpenAI's GPT-6 Astra takes the single hardest long-horizon engineering test, FrontierSWE 2, and Anthropic's Claude Opus 5.5 takes computer use, where a model has to operate a desktop interface. If your workload is driving a computer through a graphical interface, the rankings say Opus 5.5 is still the model to beat. If it is securing or rebuilding a large codebase, the rankings say Carbon.
How Carbon compares with previous Gemini models
Against Argon, the comparison is clean because both were measured on the same new index: Carbon 59 against Argon 50, with gains on every one of the 17 benchmark rows in Google's table. The security gains are large — ExploitGym up from 42.0% to 54.8%, Terminal-Bench 4.0 up from 57.1% to 65.2% — and the long-context gain is larger still. On WebDev Arena, Carbon gained 89 Elo points over Argon to reach fourth place at 1769.
Against the Gemini 3 generation, the comparison needs a warning label. Artificial Analysis changed its index between generations, so scores do not transfer directly: Gemini 3 Pro scored 39.6 on the older Intelligence Index v4.1, and Gemini 3.8 Flash scored 59 on v4.1.1 — the same number as Carbon's v5.0 score, but on a different, easier version of the test. Do not let anyone tell you the two 59s mean the models are equal; they are different rulers. What can be said fairly is that the Gemini 3 generation's public highlights — Gemini 3 Pro at 76.2% on SWE-bench Verified and 91.9% on GPQA Diamond, Gemini 3 Flash at 78% on SWE-bench Verified for just $0.50 per million input tokens — belonged to a different benchmark era, and the Gemini 4 generation has moved the contest to agentic, long-running work where raw exam scores matter less than whether the model finishes the job without drifting.
The pace itself is part of the story. Google's Q3 earnings call noted that Argon usage ramped three and a half times faster than Gemini 3 in its first week, and that internal teams shipped 35% more code with 30% fewer errors using Argon. Carbon arrived two weeks after Argon's general release. For a company that spent much of 2026 releasing smaller, efficiency-focused models while OpenAI and Anthropic shipped flagships at a furious pace, two frontier models in a month is a visible change of tempo — and analysts note the element-name pattern (Argon, Carbon) suggests a family with more siblings to come, though Google has announced none.
What it costs: pricing, and the price rise already scheduled
Carbon's API pricing matches Argon's introductory rates exactly, and that is the point: Google is pricing its new flagship like the old one to pull developers across.
| API price | Introductory rate | Standard rate |
|---|---|---|
| Input, per million tokens | $2 | $4 |
| Output, per million tokens | $10 | $20 |
| Cached input, per million tokens | 95% off (about $0.10 intro, $0.20 standard) | 95% off |
Here is the detail that separates the two models' pricing stories. Argon's introductory window has a published end: $2 and $10 run through November 30, 2026, and the standard $4 and $20 apply from December 1. Carbon's announcement carries the same $4 and $20 step-up in a footnote but does not say when its introductory period ends. Anyone planning spend around Carbon should treat the $2 and $10 rates as temporary with an unknown expiry, not as the permanent price.
On independent per-task cost analysis, Carbon costs $7.76 per Artificial Analysis Intelligence Index task at standard pricing — about 23% less than Claude Opus 5.5 at $10.04, while using 123,000 output tokens per task against Opus 5.5's 173,000, or 29% fewer. It is, however, roughly 1.2 times GPT-6 Astra's $6.25 per task, and in line with Argon's $7.68 at the same $4 and $20 rates while scoring 9 points higher on the index. For comparison, coverage notes that OpenAI's GPT-6.1 Sol and Anthropic's Claude Sonnet 5.5 sit at the same $2 input and $10 output introductory rates. In short: not the cheapest per task, but cheaper than Anthropic's flagship per unit of intelligence, and slightly dearer than OpenAI's mid-tier.
Who Gemini 4 Carbon is for — and who it is not for yet
Right now, Carbon is for a narrow group: vetted cybersecurity defenders and trusted partners admitted to Google's Fairwind Program. That is the whole first audience. If you do defensive security work — hunting vulnerabilities in real software, validating exploits, patching at scale — you are the person Google built this for, and the Fairwind route is the only door in.
Behind that first group, the natural next users are enterprise security teams, large engineering organizations with sprawling legacy codebases, and research teams doing long-horizon scientific work. Google's own earnings call already pointed at the workloads Argon proved itself on — untangling legacy codebases, accelerating scientific research, robotics, finance, and legal work — and Carbon is positioned as the stronger successor for exactly those jobs, with security as the sharp end.
Just as important is who Carbon is not for yet. It is not for everyday consumers: there is no Gemini app access, no free tier, and no plan you can buy today that includes it. It is not for developers building on the public API, because the public API does not serve it. And it is not for teams that need a model this week with a published price and a service contract — Carbon has neither a firm availability date nor a confirmed end to its introductory pricing. If any vendor, blog, or video tells you they are "using Gemini 4 Carbon" in a consumer product right now, treat that claim with skepticism: the only confirmed access is Fairwind and government testing.
Where it has launched: gated by program, not by geography
This is the question with the most misleading answers circulating, so here is the precise position. Carbon has not launched globally, and it has not launched region by region either. Its rollout is gated by program membership, not by country. Starting in the week of October 28, 2026, Google is making it available to a subset of cyber defenders, then gradually phasing access to additional trusted partners through Fairwind. Broader access for developers, businesses, and consumers comes, in Google's words, "as soon as possible" after feedback and guardrail iteration — with no date attached. Separately, Carbon is going through the United States government's voluntary pre-release model testing process, the same kind of government evaluation that now precedes several frontier releases.
The contrast with Argon shows what a real launch looks like for this family. Argon went generally available on October 13, 2026, and this is how you can actually get it today:
| Route to Argon | Who gets it | What it costs |
|---|---|---|
| Gemini app and AI Mode in Search | Google AI Pro and Ultra subscribers | Pro $19.99 per month; Ultra from $99.99 per month |
| Gemini API and AI Studio | Paid API tier only — not the free tier | $2 input / $10 output per million tokens through Nov 30, 2026; $4 / $20 from Dec 1 |
| Google Antigravity (agentic coding) | Ultra and Enterprise subscribers | Ultra from $99.99 per month |
| Vertex AI and Gemini Enterprise | Enterprise developers | Same per-token API rates; regions global, US, and EU |
| Cursor and GitHub Copilot | Third-party model pools (Copilot Pro+ and above, not base Copilot Pro) | Provider pass-through / usage billing |
| Fairwind Program | Vetted cyber defenders | Access to the variant without cyber guardrails, for defensive work |
Two groups explicitly do not get Argon: free-tier API users and Google AI Plus subscribers at $4.99 per month, who are prompted to upgrade to Pro. That tier ladder is the best available guide to how Carbon will probably be sold when it eventually opens up — paid API first, Pro and Ultra in the app, Ultra in Antigravity — but Google has confirmed none of that for Carbon itself, and this article will not present a guess as a fact.
The honest tradeoffs
Carbon's record is not a clean sweep, and the exceptions are informative. GPT-6 Astra beats it on FrontierSWE 2, the ultra-long-horizon software engineering test, 65.5% to 61.2%. Claude Opus 5.5 beats it clearly on OSWorld-2.1 computer use, 87.2% to 81.5%. And Artificial Analysis found a genuine regression hiding inside the improvement: on its factual-knowledge test, Carbon answered 58% correctly, up 10 points from Argon, but its hallucination rate rose from 4% to 11%. That is still the second-lowest rate Artificial Analysis has recorded, but the direction matters — Carbon is more accurate and more confidently wrong at the same time, and teams using it for factual work should verify outputs rather than trust the headline score.
There is also a measurement caveat that applies to the whole launch. Most rival numbers in Google's 17-row table are self-reported by each provider, Carbon was run at its highest reasoning setting, and the security evaluations ran without guardrails. Independent testing so far confirms the broad shape — Carbon first on the overall index, genuinely stronger on security and long context — but not the internal figures like the unnamed 10-times Chrome comparison. The launch also landed on a day when Anthropic's Frontier Red Team reported that an unreleased Claude had chained flaws into working browser and operating-system exploits, which is the other side of this coin: the same capabilities that make Carbon a strong defender are exactly what make frontier models a live security question, and it is why Google is gating access while it iterates on guardrails.
Do not confuse it with Genie 4
One day before Carbon, on October 27, Google DeepMind announced Genie 4 — a different product with a similar-sounding name. Genie 4 is a world model: it simulates interactive worlds with native audio in real time, and Google positions it as a building block for training robots and agents. Carbon is a language-and-reasoning frontier model for coding and cybersecurity. If you see a headline about a Google model "playing in a simulated world," that is Genie 4. If it is about finding and patching software vulnerabilities, that is Carbon.
Should you wait for Carbon, or use Argon now?
For almost everyone reading this, the practical answer is Argon now. It is generally available, it runs in the Gemini app for Pro and Ultra subscribers, it serves the paid API at the same introductory $2 and $10 rates through November 30, and its 1,048,576-token unified context window already covers the long-document and large-codebase workloads most teams actually have. Its weaknesses against Carbon are real but narrow: security scores a step lower, long-context retrieval weaker, and a lower overall index score.
Wait for Carbon if you do defensive security at serious scale — vulnerability research, exploit validation, patching across a large estate — and apply to Fairwind, because that is the only queue that exists. Everyone else should treat Carbon as a signal of where Google's frontier is heading, not as a tool to plan around: no date, no consumer plan, no public API. When Google does open it up, the price to watch is whether the $2 and $10 introductory rates survive, because the published $4 and $20 standard rates would double the cost of every long-context job overnight.
Can I use Gemini 4 Carbon today?
Almost certainly not. Access is limited to a subset of cyber defenders through Google's Fairwind Program, with gradual expansion to more trusted partners. There is no public API access, no Gemini app access, and no announced date for broader release.
What plan do I need for Gemini 4 Carbon?
No consumer or developer plan currently includes Carbon. The closest available model, Gemini 4 Argon, needs a paid API tier, or Google AI Pro ($19.99 per month) or Ultra (from $99.99 per month) for the Gemini app, with Antigravity restricted to Ultra and Enterprise. Carbon's own plan structure has not been announced.
Has Gemini 4 Carbon launched globally or region by region?
Neither. The rollout is gated by program membership, not geography: Fairwind defenders first, then more trusted partners, with broader developer, business, and consumer access promised "as soon as possible" and no date given. It is also undergoing United States government voluntary pre-release testing.
How much does Gemini 4 Carbon cost?
Introductory API pricing is $2 per million input tokens and $10 per million output tokens, with cached inputs discounted 95%. Standard pricing of $4 and $20 is already published, but unlike Argon — whose introductory rates end November 30, 2026 — Carbon's introductory end date has not been specified.
Is Carbon really better than GPT-6 Astra and Claude Opus 5.5?
On the overall Artificial Analysis Intelligence Index v5.0, yes: Carbon 59, Opus 5.5 58, GPT-6 Astra 57, and Carbon leads 15 of the 17 rows in Google's table. But GPT-6 Astra leads on FrontierSWE 2 and Opus 5.5 leads on computer use, and Carbon's hallucination rate on a factual-knowledge test rose from Argon's 4% to 11%. The honest verdict is: best overall and best at security, not best at everything.
Sources
- Google DeepMind, "Introducing Gemini 4 Carbon," Google Blog, October 28, 2026 — https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/
- Reuters, "Google launches Gemini 4 Carbon AI model with cybersecurity focus," October 28, 2026 — https://www.reuters.com/technology/google-launches-gemini-4-carbon-ai-model-with-cybersecurity-focus-2026-10-28/
- VentureBeat, "Google unveils Gemini 4 Carbon, retaking the world's no. 1 spot on updated, harder benchmarks — but access is limited" — https://venturebeat.com/technology/google-unveils-gemini-4-carbon-retaking-the-worlds-no-1-spot-on-updated-harder-benchmarks-but-access-is-limited
- Artificial Analysis, "Gemini 4 Carbon is the new #1 on the Artificial Analysis Intelligence Index v5" — https://artificialanalysis.ai/articles/gemini-4-carbon-new-number-1
- MarkTechPost, "Google DeepMind Introduces Gemini 4 Carbon, Its Top Model for Cyber Defense and Coding," October 28, 2026 — https://www.marktechpost.com/2026/10/28/google-deepmind-gemini-4-carbon-frontier-model/
- OfficeChai, "Google Launches Gemini 4 Carbon, Beats GPT-6 Astra, Claude Opus 5.5 On Most Benchmarks" — https://officechai.com/ai/gemini-4-carbon-benchmarks/
- WinBuzzer, "Gemini 4 Carbon AI Model Puts Google Back on Top, But Only Cyber Defenders Can Use It For Now," October 28, 2026 — https://winbuzzer.com/2026/10/28/gemini-4-carbon-ai-leads-test-for-fixing-vulnerable-code-xcxwbn/
- TechAIWire, "Gemini 4 Argon GA: API pricing, 1M context, availability" — https://techaiwire.com/articles/gemini-4-argon-generally-available-api-pricing/
- Dealroom, "Google DeepMind's Gemini 4 Carbon takes the #1 spot on Artificial Analysis, two weeks after Argon" — https://dealroom.co/news/167775-google-deepminds-gemini-4-carbon-takes-the-1-spot-on-artificial-analysis/
- Google, "Q3 2026 earnings call: Remarks from our CEO," Google Blog — https://blog.google/company-news/inside-google/message-ceo/alphabet-earnings-q3-2026/
Be the first to share a useful question or thought.