Two weeks ago, Google handed its strongest model to a small group of cybersecurity defenders and kept the door locked. Today the door is open, and the price tag is the interesting part.
Google announced on Tuesday, November 10, that Gemini 4 Carbon is now available to developers, enterprises, scientists and consumers. Carbon had been out since late October, but only inside the Fairwind Program for cyber defenders. Now it is in the Gemini API for every paying developer, in Gemini for business, and rolling out to Google AI Ultra subscribers in Google Antigravity, the Gemini app and AI Mode in Search. The flagship comparison on this site just changed shape again, one day after Anthropic rewrote it.
What Carbon is, in Google’s own framing
Google calls Carbon its strongest model, built for frontier cybersecurity work and long horizon agentic tasks across knowledge work, software engineering and scientific reasoning. The division of labor in the Gemini 4 family is explicit in the post: Argon remains the pick for most software engineering and knowledge work, while Carbon is for the jobs that test the limits, complex multi step coding and debugging, open ended scientific research, the tasks where a model needs to spend more tokens, reason longer and run extended verification loops before it answers.
The launch also previews a feature called /swarm in Google Antigravity. One prompt becomes a coordinated team of agents that you can track in real time, with Carbon assigned to the deep reasoning roles like architecture, coordination and auditing, and faster, lighter models working in parallel workspaces with adversarial verifiers checking every change. That is Google describing its own product, so treat the framing as a pitch. The feature itself is real and in preview.
The price is the headline
Carbon costs the same introductory price as Gemini 4 Argon: 2 dollars per million input tokens and 10 dollars per million output tokens, per Google’s post.
Put that next to yesterday’s news. Anthropic launched Claude Fable 5.2 on Monday at 8 dollars per million input tokens and 40 dollars per million output, and that was a 20 percent price cut that undercut GPT-6 Astra on input. Carbon lands at exactly a quarter of Fable’s price on both sides of the meter. Price is not capability, and the two models are aimed at overlapping but not identical work. Still, when the newest frontier model in the comparison costs a quarter of the one launched the day before, every developer doing the math this week just felt the floor move.
The benchmark claims, labeled as claims
Everything in this section is Google reporting its own results, and I am keeping the label on.
Google says Carbon ranks number one on CursorBench 4.0, on Perplexity’s WANDR benchmark and on Zapier’s AutomationBench, at more than twice lower cost per task than the next best external model. On a new Multimodal Knowledge Agent benchmark built with Artificial Analysis, Google says Carbon ranks second while costing more than six times less per task than the top model. Inside Google’s own Project Fortify security system, the company reports Carbon hit 100 percent recall on scanning and 95 percent precision on triage on an internal benchmark, and improved open source vulnerability recall by more than 25 percent relative to Argon. On safeguards, Google says Carbon blocked 99 percent of attack trajectories from its adaptive red team agent in simulated testing, cut the per attempt breach rate to 0.8 percent from Argon’s 3.1 percent, and reached a 0 percent prompt injection success rate on the Gray Swan benchmark, which the company calls a first.
Those are strong numbers if they hold up outside Google’s own testing, and that conditional is doing real work in this paragraph. Independent confirmation is the next story here, the same way it was for Anthropic’s black hole claims yesterday.
My take
The AI race this fall has stopped being about who has the smartest model for more than about a week at a time. It is about who can sell frontier reasoning cheaply enough that nobody bothers downgrading. Anthropic cut prices on Monday and looked aggressive doing it. Google answered on Tuesday at a quarter of the price, with the cyber specialist model it had been keeping behind a velvet rope for defenders only.
For readers of the comparison piece on this site, the practical read is simple. If your work is long documents and careful writing, Claude still has its case. If your life runs through Google’s stack and your agents need to grind through big, ugly, multi step jobs, Carbon just made itself very hard to ignore at 2 and 10 dollars. The subscription is still 20 dollars a month either way. What changed is the meter running underneath it, and meters are where the real money moves.
Sources
Google, “Gemini 4 Carbon is now available”, official Google blog, November 10, 2026
Google, Gemini 4 Carbon launch coverage via The Neuron daily roundup, November 10, 2026
Anthropic, Claude Fable 5.2 and Mythos 5.2 launch, November 9, 2026 (published on this site, post 2316)
The Checkpoint Poll
Leave a Reply