Google has opened Gemini 4 Carbon, its strongest AI model yet, to paying developers and top-tier subscribers. It costs a fraction of what rivals charge, and it comes with a new tool that turns one prompt into a whole team of AI agents.
Google made Gemini 4 Carbon generally available on Tuesday, November 10, two weeks after first releasing it only to a select group of cybersecurity defenders. The company calls Carbon its "strongest model" for cybersecurity and for long, multi-step agent work across software engineering, knowledge work and scientific reasoning, according to Google's announcement by Google DeepMind's Koray Kavukcuoglu. Alongside the launch, Google is previewing a feature called /swarm in its Antigravity coding platform, which coordinates several AI agents working on one project at the same time.
What Google Released
Carbon is now available through the paid Gemini API in Google AI Studio and Android Studio, and in Gemini for business for enterprises. It is also rolling out to all Google AI Ultra subscribers in the Gemini app, AI Mode in Search and Antigravity. Google's US Ultra plans cost $99.99 a month for Ultra 5x and $199.99 a month for Ultra 20x, VentureBeat reports. The Verge notes that it is not clear when, or if, Carbon will reach lower-tier subscribers or free users.
Google is not replacing its other frontier model. It says Gemini 4 Argon, launched last month, "remains the ideal choice for the majority of software engineering and knowledge work." Carbon is aimed at harder jobs, such as complex coding and debugging or open-ended research, and it uses more tokens than Argon to "explore thoroughly, reason deeply, and run extended verification loops." A more capable version, Carbon for cyber, stays behind Google's gated Fairwind Program for vetted security teams.
The Price and the Numbers
Carbon launches at an introductory price of $2 per million input tokens and $10 per million output tokens, the same as Argon. That rises to $4 and $20 after the promotion ends on November 30. The Verge says the launch price is about a fifth of the cost of OpenAI's GPT-6 Astra and a quarter of Anthropic's Claude Fable 5.2, which arrived a day earlier.
Google says Carbon ranks first on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench, at "more than 2X lower cost per task" than the next-best outside model. Independent data supports part of that claim. On Zapier's AutomationBench, Carbon scored 53.88% at about $2.12 per task, ahead of Fable 5.2 at 51.45% and $4.64, VentureBeat found. But VentureBeat also said the public CursorBench and WANDR tables it reviewed did not include Carbon, so Google's claims on those two tests remain unverified.
On Artificial Analysis's broad intelligence index, Carbon debuted in first place in October, but Fable 5.2 now leads with 65 points against Carbon's 60. Coding results are split. Carbon leads on Terminal-Bench 4.0 with 65.2% versus Fable's 61.6%, while Fable leads on FrontierCode 2.0, a test of real repository changes, with 49.1% against Carbon's 40.6%.
How /swarm Works
The new /swarm tool takes a single prompt and hands it to a coordinator agent, which starts specialised workers in separate git worktrees so their edits do not collide. Dedicated "adversarial verifiers" run tests and challenge each change, and all agents post to a shared message board, according to the Antigravity team. Users can assign different models to different roles, putting Carbon on architecture, coordination and auditing while lighter models handle routine edits. Google says its engineers tested /swarm internally for five weeks on large migrations and debugging, and it warns that multi-agent runs "consume quota much faster than a standard chat."
Cybersecurity, Science and Safety Claims
Google says Carbon leads key cyber benchmarks, scoring 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for exploiting them and 71.0% on CWE-bench for patching them. It also says cloud security firm Wiz used Carbon to uncover previously unknown "0-day" flaws at two major software-as-a-service providers in under two hours. Google did not name the providers.
In research, Google says its AlphaProof Code system used Carbon to help produce millions of lines of Lean proofs verifying security properties of Signal's encryption code in weeks, work it says has historically taken expert teams years. On safety, Google reports that Carbon blocked 99% of simulated attack attempts by an automated red-teaming agent, and that it passed 92% of a 375-task honesty test, up from 70% for Argon. These figures come from Google and have not been independently confirmed.
Why It Matters
The Verge calls the release "a decisive return for Google at the very forefront of AI research" after months in which frontier AI looked like "a two-horse race between Anthropic and OpenAI." It also cautions that companies "can and do cherry-pick benchmark results."
Our analysis: Carbon's real pitch is cost, not a clear win on raw ability. It trails Fable 5.2 on the broad index but undercuts it sharply on price per token. As VentureBeat points out, however, a cheaper token does not always mean a cheaper job, because models use different amounts of tokens, retries and human fixes to finish the same task.
What Comes Next
The key date is November 30, when Carbon's price doubles. Teams testing it now will learn whether its extra capability is worth paying for at full rates. Google is also expanding a programme that gives selected scientists six months of free access to Antigravity's highest tier, now covering Carbon as well as Argon. With Anthropic, OpenAI and Google releasing new models within days of each other, buyers may soon choose AI models the way /swarm does: by matching each model to the job it does best.