Google’s strongest AI model is no longer reserved for cybersecurity teams. On November 10, Google opened Gemini 4 Carbon to paying developers, enterprises, scientists and consumers, two weeks after announcing it to a small group of cyber defenders. It arrives with three benchmark first places, a new way to run teams of agents, and a price that doubles on December 1.
Carbon is the heavyweight of the Gemini 4 family. Google says it is built for tasks that test the limits of a model’s intelligence, like complex multi-step coding, debugging and open-ended scientific research, and it uses more tokens than its sibling Argon to explore thoroughly, reason deeply and run extended verification loops. The question for buyers is not whether it is smart. It is whether the extra tokens are worth it, and the answer depends on how you count the cost.
From cyber defenders only to everyone in two weeks
Carbon was announced on October 28, 2026, but almost nobody could use it. Google gave it first to a subset of cybersecurity defenders through its Fairwind Program, with wider access promised later. The November 10 post from Koray Kavukcuoglu, SVP of Google DeepMind, ends that wait. Carbon is now available through the Gemini API paid tier in AI Studio and Android Studio, and in Gemini for business. Google AI Ultra subscribers also get it in Antigravity, the Gemini app and AI Mode in Search. There is no free API tier.
Scientists get a separate path. Google is expanding its research program to give selected participants both Argon and Carbon in Antigravity’s highest usage tier free for six months. The company says it opened applications for 10,000 scientists four weeks ago and has already onboarded thousands.
The benchmark table: three first places, one caveat
Google says Carbon ranks first on CursorBench 4.0, Perplexity’s WANDR and Zapier’s AutomationBench, at more than twice lower cost per task than the next best external model. The AutomationBench numbers, on 657 tasks at High effort, are specific enough to check:
| Model | AutomationBench 1.0.6 | Intro price per million tokens |
|---|---|---|
| Gemini 4 Carbon | 53.9% | $2 input / $10 output |
| Gemini 4 Argon | 51.3% | $2 input / $10 output |
| Claude Sonnet 5.5 | 44.7% | Not stated by Google |
| GPT-6 Astra | 41.4% | Not stated by Google |
Google also launched a new benchmark with Artificial Analysis, the Multimodal Knowledge Agent test, where Carbon ranks second in quality at more than six times lower average cost per task than the top model. On Artificial Analysis’s Intelligence Index v5, Carbon scores 59, against 50 for Argon.
The caveat is that every one of these figures comes from Google. There are no independent measurements yet. Carbon’s lead over Argon on AutomationBench is about 17 tasks out of 657, which could sit inside the margin of noise, and Carbon was tested at High effort while the comparison of effort and compute with rivals is not described. Treat the rankings as a claim to watch, not a verdict.
Carbon vs Fable 5.2: the price per task matters more
The more interesting comparison is with the model that currently tops the Intelligence Index. Anthropic’s Claude Fable 5.2, released November 9, scores 64.6, ahead of Carbon’s 59, and costs $8 per million input tokens and $40 per million output, four times Carbon’s introductory rate. Artificial Analysis puts Carbon at $7.76 per Intelligence Index task at standard pricing, about 23 percent less than Claude Opus 5.5 at maximum effort.
That framing is the point Google wants buyers to adopt. Agentic work burns an order of magnitude more tokens than a single chat request, so the cost of a solved task matters more than the cost of a token. Carbon’s extra reasoning tokens are billable, including thinking tokens, so matching Argon’s per-token price does not guarantee matching per-task costs. The honest way to compare is to run the same pipeline through both models and count what each solved task actually cost. The introductory price makes that experiment cheap for the next twenty days.
The November 30 deadline on the intro price
Carbon launches at the same introductory price as Argon, $2 per million input tokens and $10 per million output, with cached input at $0.10. That rate ends November 30, 2026. From December 1 the standard price is $4 per million input tokens and $20 per million output, double the intro rate. Anyone building a permanent pipeline on Carbon should set up routing and caching before the deadline, or the bill doubles without warning.
The open-weight world offers a cheaper route for some workloads. Moonshot’s Kimi K3.1 Pro, whose weights are free to download, costs $2 per million input tokens and $10 per million output on its API, the same as Carbon’s intro rate, and scores 46.5 on the Intelligence Index. It is not in Carbon’s class on agentic benchmarks, but for routine work it is a reminder that the frontier models now compete on cost per solved task, not just on raw scores.
Antigravity /swarm turns one prompt into a team
Alongside Carbon, Google is previewing /swarm in Antigravity. A single prompt launches a coordinated team of agents that write, check and test code in parallel, with Carbon assigned to the roles that demand the deepest reasoning, like architectural design, coordination and auditing, while faster, lighter models handle the rest across parallel isolated workspaces. Adversarial verifiers test every change. Google has not said when /swarm leaves preview or whether it will get API primitives.
The cybersecurity case is where Carbon looks most distinct. On Google’s own benchmarks it scores 88.2 percent on SEC-bench Pro for finding vulnerabilities, 54.8 percent on ExploitGym for checking exploitability and 71.0 percent on CWE-bench for patching. In Wiz’s Scan for Good program, the model found zero-day vulnerabilities in two major SaaS providers in under two hours, work Google says would normally take weeks. Its Gray Swan prompt injection attack success rate is 0 percent, the first time that milestone has been reached, though Google admits robustness against sophisticated attacks still needs work.
The science results are striking but early. For drug toxicity research company Axiom Bio, Carbon is now the leading model for translational science, finding toxicity mechanisms the team had not found before. Working in a multi-agent setup, Google researchers used Carbon to discover 40 times more protein homologs on average across seven low-homolog families from the CASP17 blind test set. That result has not yet been reproduced by outside teams.
Carbon is the strongest Gemini yet, and for the next twenty days it is also the cheapest time to test it. The November 30 price rise is the real deadline. Run a real pipeline through it before then, count the cost per solved task, and decide with your own numbers rather than Google’s benchmarks. Google’s new TPU 8tx chip, which locks sensitive AI inside a secure enclave, and TSMC’s $500 million investments in Anthropic and OpenAI show the same race from the hardware side. The models are getting cheaper to run. The question is which one solves your task for less.
Sources: Google Blog, Zapier Benchmarks, Artificial Analysis, n8n, Unite.AI.
