Takeaway: On November 10, 2026, Google said Gemini 4 Carbon is now available to developers, enterprises, scientists, and consumers — with important limits on each path. Paying developers and enterprises can use the model through the Gemini API in Google AI Studio and in Gemini for business, at the same introductory price as Gemini 4 Argon: $2 per million input tokens and $10 per million output tokens. Google AI Ultra subscribers also get Carbon in Google Antigravity, the Gemini app, and AI Mode in Search. After November 30, 2026, that public-access Carbon model moves to $4 / $20, according to the same post. This is the general availability follow-up to our October 28 coverage of the Fairwind-limited announcement, when Google had not set a general-availability date.
Two things did not change. Google still says Argon “remains the ideal choice for the majority of software engineering and knowledge work,” while Carbon is for harder, longer jobs that use more tokens. And “Carbon for cyber” — the cybersecurity configuration Google discusses with SEC-bench Pro, ExploitGym, and CWE-bench — still goes through the Fairwind Program, not automatically through the new API and Ultra surfaces. Google has not said when, or whether, free or lower-tier Gemini and AI Mode users will get Carbon.
What changed on November 10
The October 28 announcement introduced Carbon as Google’s strongest model for cybersecurity and long-horizon work, then described a deliberate rollout: U.S. voluntary pre-release testing, access first for a subset of cyber defenders, and a gradual expansion to trusted partners through Fairwind. Broader product access was promised “as soon as possible,” without a date. Koray Kavukcuoglu’s November 10 post is the date. It says Carbon was released to a set of cybersecurity defenders via Fairwind two weeks earlier, and that Google is now making the model available across the product family listed above.
A companion developers post repeats the practical starting points: the paid tier of the API in Google AI Studio and Android Studio, plus Antigravity and AI Studio for Google AI Ultra subscribers. Enterprises can try Carbon in Gemini for business through the API or a model dropdown, in Antigravity, or in CodeMender. Scientists and researchers can still apply for six months of Antigravity’s highest usage tier, now including Carbon as well as Argon, subject to eligibility, capacity, and selection, with weekly rolling onboarding.
Treat “available to consumers” as Ultra-gated in the Gemini app and in AI Mode, not as a default switch for every Google account. AI Mode itself is a separate Search surface; Carbon’s arrival there is a model option for Ultra subscribers, not a ranking or ads change.
Pricing, and what the footnote covers
The introductory price is $2 per million input tokens and $10 per million output tokens, the same intro rate Google used for Argon. The November 10 post is explicit about the end of that window: after November 30, 2026, $4 per million input tokens and $20 per million output tokens apply. A footnote says that later rate “applies to the Carbon model intended for public access only.” Cached input pricing from the October announcement — 95% off the input token price — was not restated in the sentences we are relying on for the November 30 change, so confirm the current API price list before you forecast a production bill.
Google also argues that Carbon can be cheaper per completed task even when it uses more tokens than Argon, because it ranks first on several workflow benchmarks “at more than 2X lower cost per task compared to the next best external model.” That is Google’s claim, tied to CursorBench 4.0, Perplexity’s WANDR, and Zapier’s AutomationBench. We have not independently rerun those evals. Cost per task is not the same thing as cost per million tokens, and it will not hold if your workload is short, shallow, or a poor fit for Carbon.
Who should start with Carbon, and who should stay on Argon
Google’s own split is the one to start from. Argon is still positioned for most software engineering and knowledge work. Carbon is “built for tasks that test the limits of a model’s intelligence,” such as complex, multi-step coding and debugging and open-ended scientific research. The company says Carbon uses more tokens than Argon to explore, reason, and run extended verification loops. If a job already finishes well on Argon, moving it to Carbon is a cost and latency decision, not an automatic upgrade.
The November post leans hard on long-running demos in Google Antigravity, including a preview of /swarm, which turns a single prompt into a coordinated team of agents you can track in real time. Google’s suggested pattern is to give Carbon the deepest roles — architectural design, coordination, and auditing — and to pair it with faster, lighter models in parallel workspaces, plus “adversarial verifiers” that test each change. Those demos are company illustrations: a pelican-and-GBike campus game, an Eiffel Tower exhibition from historical plates, a Platform 37 office reconstruction, and other Three.js and Blender projects. They show what Google wants Carbon to be used for. They are not a substitute for a test on your own codebase.
New benchmark claims — and the caveats from October
Alongside the three first-place workflow leaderboards named above, Google says it partnered with Artificial Analysis on a new Multimodal Knowledge Agent (MKA) benchmark for agentic work over dense files with charts, images, blueprints, and tables. On that test, Google says Carbon ranks second while coming in at more than 6X lower average cost per task than the top model. It also reports strong results on Hebbia’s end-to-end financial agentic evaluation and a top-of-field result on Balyasny Asset Management’s General Financial Intelligence benchmark, across 23 frontier model configurations. Those figures are from Google’s November 10 post. We have not audited the harnesses or the cost methodology.
Our October 28 article drew a sharper line on the cybersecurity scores Google published at announcement: CWE-bench was checkable on a public leaderboard, while SEC-bench Pro and ExploitGym figures were Google-reported. That distinction still matters. The November post repeats that Carbon leads on finding, exploiting, and patching benchmarks “when accessed through the Fairwind Program.” Do not read a new API key as automatic access to that cyber configuration, and do not treat a Google-reported percentage as a public board unless you can open the board yourself.
Cyber access stays on Fairwind
The November post describes early results from Project Fortify, Google’s Cloud Vulnerability Research team, and Wiz’s Scan for Good program, including a claim that Carbon hit 100% recall on scanning and 95% precision on triage in an internal Fortify benchmark, and that Wiz used the model to find 0-day vulnerabilities in under two hours at two major SaaS providers. Those are impressive anecdotes from named teams. They are also Google-selected examples, and some details are held back for responsible disclosure. They do not change the access rule: if you want Carbon for cyber, Google still says to apply to Fairwind.
Safety claims are likewise Google’s. The company says Carbon blocked 99% of attack trajectories in simulated environments seeded with ExploitBench and CyScenarioBench, lowered a per-attempt breach rate to 0.8% from 3.1% on Argon, and, as it already reported in October, reached a 0% prompt-injection success rate on the Gray Swan IPI benchmark — while also warning that prompt-injection robustness still needs work against more sophisticated attacks. On an internal 375-task agentic honesty benchmark, Google says Carbon passed 92% of tasks, compared with 70% for Argon. Read the company’s safety and security report if those numbers are part of a procurement review; we are not restating them as independent measurements.
How this sits next to other Google news this week
Carbon’s public rollout is a model-access story. It is separate from Google’s November 9 TPU 8tx and 8i Max announcement, which added confidential-inference and dense-context systems to the eighth-generation TPU family, and from Maps Platform inside Google AI Studio, a prototyping path with its own demo-key limits. You can use Carbon in AI Studio without those hardware or Maps launches being part of the same product.
Access and price at a glance
| Path | What Google’s November 10 posts say |
|---|---|
| Paying developers | Gemini API paid tier in Google AI Studio and Android Studio. |
| Google AI Ultra subscribers | Google Antigravity, AI Studio, the Gemini app, and AI Mode in Search. |
| Enterprises | Gemini for business via the API or model dropdown, Google Antigravity, or CodeMender. |
| Scientists and researchers | Apply for Argon and Carbon on Antigravity’s highest usage tier, free for six months, onboarded on a rolling basis, subject to eligibility, capacity, and selection. |
| Cybersecurity defenders | Apply to the Fairwind Program for Carbon for cyber. This is still a separate path. |
| Free or lower-tier consumer plans | Not listed in the November 10 access list. Google has not announced a date for those tiers. |
| Introductory API price | $2 per million input tokens and $10 per million output tokens, through November 30, 2026. |
| Price after November 30 | $4 per million input tokens and $20 per million output tokens, for the Carbon model intended for public access. |
FAQ
Is Gemini 4 Carbon generally available now?
It is available on the paths Google listed on November 10, 2026: the paid Gemini API, Gemini for business, and Google AI Ultra surfaces including Antigravity, the Gemini app, and AI Mode in Search. That is a major expansion from the October 28 Fairwind-first rollout. It is not the same as Carbon being on by default for every Gemini or Search user.
Should I replace Gemini 4 Argon with Carbon?
Not automatically. Google still recommends Argon for most software engineering and knowledge work, and says Carbon uses more tokens on harder, longer tasks. Test Carbon on the jobs that actually stall on Argon, and compare quality, latency, and cost per task before you migrate a production default.
When does the $2 / $10 price end?
Google says the introductory period expires on November 30, 2026, after which $4 per million input tokens and $20 per million output tokens apply to the Carbon model intended for public access. Confirm the live API price page when you budget, including any cached-input rates.
Is this the same announcement as October 28?
No. October 28 was the model announcement and a limited Fairwind rollout, with no general-availability date. November 10 is the post that opens the paid API and named product surfaces. Our earlier article still applies for the original benchmark caveats and the Fairwind framing.
What is /swarm?
Google is previewing /swarm in Google Antigravity. It turns a single prompt into a coordinated team of agents, with Carbon suggested for the deepest reasoning roles and lighter models plus verifiers running in parallel. It is a preview, not a claim that every Carbon API call now spawns a swarm.
Does API access include Carbon’s cybersecurity mode?
Google discusses its strongest cyber results, and the SEC-bench Pro, ExploitGym, and CWE-bench lead, in the context of Carbon accessed through the Fairwind Program. The November 10 getting-started list still sends cybersecurity defenders to Fairwind. Do not assume a standard API key is that configuration.
Can I use Carbon in AI Mode or the Gemini app without an Ultra plan?
Google’s November 10 consumer path is explicit: try Carbon in the Gemini app and in AI Mode in Search if you are a Google AI Ultra subscriber. The posts do not list a free or lower-tier consumer rollout date.
Is this related to Google’s new TPUs or to the November core update?
No. TPU 8tx and 8i Max are infrastructure announcements from November 9. The November 2026 Search core update is a ranking-systems rollout. Carbon’s November 10 post is about model access, pricing, and product surfaces.