Gemini 4 Carbon is out for Google AI Ultra and the API. Who gets it, and what it costs

What changed on November 10

On November 10, 2026, Google opened Gemini 4 Carbon to the public: “Today, we’re making Gemini 4 Carbon available to developers, enterprises, scientists, and consumers.” Google calls it “our strongest model,” with “frontier performance in cybersecurity capabilities and long-horizon agentic workflows.”

When this memo first went out, on October 28, Google had announced Carbon but given it only to “a subset of cyber defenders” through its Fairwind Program, and said it would reach everyone else “as soon as possible.” Two weeks later, it has, though not on every plan.

Its model card says “Gemini 4 Carbon is based on Gemini 4 Argon,” and Google is clear that it doesn’t replace Argon: “Argon remains the ideal choice for the majority of software engineering and knowledge work, while Carbon is built for tasks that test the limits of a model’s intelligence like complex, multi-step coding and debugging and open-ended scientific research and discovery — using more tokens than Argon to explore thoroughly, reason deeply, and run extended verification loops.”

Who gets it, and where

In Google’s own apps, Carbon is for Google AI Ultra. The announcement says it is “rolling out to all Google AI Ultra subscribers in Google Antigravity, the Gemini app, and AI Mode in Search,” and that Ultra subscribers can also try it in AI Studio. Developers get it through the Gemini API’s paid tier, and businesses through Gemini for business.

Here is Carbon beside Argon, plan by plan, as Google’s pages read on November 10:

Plan or account	Price	Gemini 4 Carbon	Gemini 4 Argon
Free and Google AI Plus	$0, or $4.99 a month	No	No, in Antigravity
Google AI Pro	$19.99 a month	No	Yes, on a paid plan
Google AI Ultra	$99.99 a month, or $199.99 for higher limits	Yes: Gemini app, Antigravity, AI Mode, AI Studio	Yes
Gemini Enterprise	Per seat	Yes, in Antigravity and Gemini for business	Yes
Gemini API, paid tier	Per token	Yes, as gemini-4-carbon	Yes, as gemini-4-argon
Gemini API, free tier	Free	No	No

The plans are from Antigravity’s models page, which marks Carbon for Ultra and Enterprise only, and Google’s AI plans, which list “Highest access to Gemini 4 Carbon” under Ultra and Argon under Pro. Ultra’s $99.99 tier has “5x higher usage limits vs. AI Pro” and its $199.99 tier 20 times. The API’s pricing page says both Gemini 4 models are “not available on the Free tier.” In Antigravity, Carbon works in the app, the CLI and the IDE extensions, Google’s Antigravity post says.

Two more groups get it free or by application. Scientists can apply for both Argon and Carbon “via Google Antigravity’s highest usage tier free of charge for 6 months,” onboarded weekly, “subject to eligibility, capacity, and selection.” Cyber defenders can apply to the Fairwind Program for “Carbon for cyber”; Google says that “when accessed through the Fairwind Program,” Carbon leads its tests of finding, exploiting and patching vulnerabilities.

“Rolling out” means an Ultra account may not see Carbon on the first day. In the Gemini app, 9to5Google reports that the model picker labels Argon “Code & analysis” and Carbon “Most complex work,” and that “the Gemini app has removed 3.1 Pro with this launch.” The app’s limits are compute-based and refresh “every 5 hours until you reach your weekly limit,” and depend on “the complexity of your prompt, the models and features you use, and the length of your chat.”

What it costs

On the API, Carbon costs the same as Argon, and both double on December 1, per Google’s pricing page, in US dollars per million tokens:

Tier	Input	Output, thinking included	Cached input
Standard, through November 30	$2	$10	$0.10
Standard, from December 1	$4	$20	$0.20
Batch or Flex, through November 30	$1	$5	$0.05
Batch or Flex, from December 1	$2	$10	$0.10
Priority, through November 30	$3.60	$18	$0.18
Priority, from December 1	$7.20	$36	$0.36

The same price per token isn’t the same price per task. Google’s developer guide says Carbon “spends more thinking tokens per turn” than Argon, “which increases per-turn token use and latency.” Its default thinking level is medium, of low, medium and high, and its context window is 1,048,576 tokens, shared by input, thinking and output. Explicit context caching “is not supported on Gemini 4 Carbon”; implicit caching is on by default for prompts of at least 4,096 tokens.

In Antigravity, Carbon draws on your plan’s quota, which Google doesn’t size. Ultra gets “The highest, most generous quota, refreshed every five hours,” per the plans page, and Pro and Ultra can buy AI credits for use beyond it, “consumed at standard Gemini Enterprise consumption pricing.”

How it compares

On Artificial Analysis’s leaderboard, now on version 5.0.1 of its Intelligence Index, Carbon is no longer first: Claude Fable 5.2, which Anthropic released on November 9, passed it. Here are the leading models, each at its highest effort, with what a task cost Artificial Analysis at API prices, on November 10:

Model	Index	Cost per task	API price per million tokens, in and out	Where you can use it
Claude Fable 5.2 (max)	65	$26.54	$8 and $40	Claude Max and Ultra; Pro through November 16
Gemini 4 Carbon (high)	60	$7.76	$2 and $10, then $4 and $20	Google AI Ultra
Claude Opus 5.5 (max)	59	$10.04	$4 and $20	Every paid Claude plan, from $20 a month
GPT-6 Astra (max)	57	$6.24	$10 and $50	ChatGPT Plus ($20 a month) and Pro
GPT-6.1 Sol (max)	55	$1.27	$2 and $10	ChatGPT Plus and Pro
Gemini 4 Argon (high)	50	$9.95	$2 and $10, then $4 and $20	Google AI Pro and Ultra

The prices are from Anthropic, OpenAI and Google, and the plans from Claude’s pricing, ChatGPT’s pricing and Google’s pages above. Of the six, Carbon gives the second-highest score at the third-lowest cost per task.

Google’s own benchmark table compares Carbon with Argon, GPT-6 Astra, Claude Opus 5.5 and Claude Fable 5.1 on 17 results, and Carbon has the best score of the five on 14. GPT-6 Astra leads FrontierSWE 2, a coding test, at 65.5% to Carbon’s 61.2%; Opus 5.5 leads OSWorld-2.1, computer use, at 87.2% to 81.5%; and Argon leads RiemannBench, a math test, at 76.0% to 72.0%. The table has Fable 5.1, not Fable 5.2. Testing on its own, Vals AI ranks Carbon third of 36 models on its Vals Index, at 70.45%.

For coding

Artificial Analysis’s Coding Agent Index scores an agent and its model together, and it now lists Antigravity with Carbon as available. The leaders, and Argon in the same agent, on November 10:

Agent and model	Index	Cost per task	Time per task
Claude Code, Opus 5.5 (max)	61.7	$11.42	61 min
Devin Fusion CLI, Opus 5.5 with GPT-6 Luna	61.5	$8.00	51 min
GitHub Copilot, Copilot Fusion (max)	59.9	$5.59	36 min
Antigravity CLI, Gemini 4 Carbon (high)	59.4	$7.74	30 min
Claude Code, Sonnet 5.5 (max)	59.3	$13.30	80 min
Codex, GPT-6 Astra (max)	54.7	$7.04	25 min
Antigravity CLI, Gemini 4 Argon (high)	53.2	$11.73	35 min

Among single-model setups, only Claude Code with Opus 5.5 scores higher than Antigravity with Carbon, and Carbon’s tasks cost about two-thirds as much and take half the time. Against Argon in the same agent, Carbon scores 6.2 points more, for about two-thirds of the cost per task. The index hasn’t tested Claude Fable 5.2 in an agent yet.

/swarm, the feature that came with it

Antigravity shipped Carbon with /swarm, “rolling out today in preview across the Antigravity App and CLI.” Type /swarm and a task, and your conversation becomes a coordinator: it splits the work, adds a named worker for each piece, and usually adds a @verifier “for every two to four workers” that “reads the diffs, runs the tests, and reports problems.” Everyone posts to one shared board, and an idle member “goes to sleep on the board without using tokens.”

Workers “run on the same model as your conversation” unless you ask for another, which “must be one you can choose in the model picker.” That’s where Carbon comes in: Google suggests it for “coordinator, architecture, cybersecurity auditing, and adversarial verification roles while lighter, faster models handle the parallel file edits.” Google’s post doesn’t limit /swarm to Ultra, but a swarm with Carbon in it needs a plan that has Carbon.

Mind the quota. Google’s warning: “parallel multi-agent systems consume quota much faster than a standard chat.” A member that runs out pauses, and “doesn’t restart on its own when your quota refreshes.”

What Google says about safety

Google’s November 10 post makes these claims, from its own and partners’ testing:

• Misuse. In simulated cyberattacks, Carbon “blocked 99% of attack trajectories,” and its safeguards “lowered the per-attempt breach rate to just 0.8% (down from 3.1% on Argon).”
• Jailbreaks. Trajectory Labs “found zero new universal jailbreaks across 132 hours and over 13,000 requests.”
• Honesty. On a 375-task agentic honesty test, “it passed 92% of tasks (compared to 70% for Argon).”
• Prompt injection. Google repeats Carbon’s 0% attack success rate on Gray Swan’s benchmark, and adds that “prompt injection robustness remains an area for continued improvement.”

Google has now published a safety and security report for Carbon itself.

What to do now

1. On Google AI Ultra, try Carbon on your hardest job. A long debugging session or an architecture question is what it’s for. Keep Argon for everyday work, as Google advises, and watch your quota.
2. On Google AI Pro, don’t upgrade for Carbon alone. Ultra is five times Pro’s price. Pro keeps Argon, and in Antigravity Claude Sonnet and Opus 5.5 too; our Pro vs Ultra guide weighs the rest. If you need Carbon for only a few tasks, the paid API in AI Studio bills per token, at the introductory price through November 30.
3. If you build on the API, budget for December 1. The price doubles that day. Google’s guide also says to raise a max_output_tokens below 32K to at least 32,768, since Carbon “can spend more of the limit on thinking,” and to remove explicit caching calls.
4. If you’re a scientist or a cyber defender, apply. The research program is free for six months, and both admit people on a rolling basis.

What Google hasn’t said

• Whether Carbon comes to Google AI Pro. Nothing on Pro, Plus or the free tier.
• How much quota it uses. Google doesn’t size Antigravity’s quota or say how much faster Carbon spends it than Argon.
• When /swarm leaves preview, or whether it will be limited to some plans.
• How Carbon differs from Argon underneath. The model card says only that it “is based on Gemini 4 Argon,” and sends readers to Argon’s card for its architecture and training data.

Carbon in Enjoy

Enjoy runs Antigravity on your Google account, through the server Google offers to apps such as Zed and JetBrains IDEs, beside Claude Code, Codex and the other labs’ agents, each on its own plan. Its model list shows “the Gemini models Google offers your account,” and Enjoy asks the agent for its current models each time you open its settings, so if Google offers Carbon to your account there, it appears in the model list without an Enjoy update. /swarm is in Antigravity’s own app and CLI; Google doesn’t list it for other apps.

That makes it easy to give a hard bug to Carbon, then switch agents in the middle of the conversation to have Claude Code with Opus 5.5 check the fix. Enjoy is a monthly subscription, with the first seven days free (pricing). It doesn’t add Carbon to an account Google hasn’t given it to.

We’ll update this memo if Google brings Carbon to Google AI Pro, or when /swarm leaves preview.