Summary: Google DeepMind made Gemini 4 Carbon available to all paying developers and enterprises on November 10, 2026, through the API in Google AI Studio and Gemini for business, at the same introductory price as Gemini 4 Argon: $2 per million input tokens and $10 per million output tokens. Carbon is the heavier sibling — it spends more tokens per task on exploration and verification — so the real question for builders is not whether it is capable, but whether that burn stays cheaper than the alternatives on your workload.
What changed today
Three things, all usable now rather than on a roadmap:
- API access. Carbon is live in Google AI Studio and Gemini for business for all paying developers and enterprises. It is also rolling out to Google AI Ultra subscribers in Google Antigravity, the Gemini app, and AI Mode in Search.
- Price parity with Argon. The introductory API price matches Argon exactly at $2/$10 per million input/output tokens. Google labels it introductory, which is worth remembering when you model costs past the intro period.
- A /swarm preview. In Google Antigravity, /swarm turns a single prompt into a coordinated agent team you can track in real time. Carbon takes the deepest-reasoning roles — architectural design, coordination, auditing — while faster, lighter models run across parallel isolated workspaces, with adversarial verifiers testing every change.
We covered the opening of availability itself in Gemini 4 Carbon opens up to all builders at Argon pricing; this post is about what the pricing and the benchmark claims mean for adoption decisions.
How Carbon differs from Argon
Google’s own positioning is explicit: Argon remains the ideal choice for the majority of software engineering and knowledge work. Carbon is built for tasks that test the limits of a model’s intelligence — complex, multi-step coding and debugging, and open-ended scientific research — and it gets there by using more tokens than Argon to explore thoroughly, reason deeply, and run extended verification loops.
That mechanism is why per-token parity is not per-task parity. Two models billed at the same $2/$10 can land at very different costs per finished task if one runs three verification passes where the other runs one. The price tag is identical; the bill is not.
The cost-per-task claims, and who measured them
Google reports that Carbon ranks #1 on CursorBench 4.0, Perplexity’s WANDR, and Zapier’s AutomationBench at more than 2X lower cost per task than the next best external model — benchmarks Google describes as independent tests of real workflows. On the Multimodal Knowledge Agent benchmark, which Google co-launched with Artificial Analysis, Carbon ranks #2 while costing more than 6X less per task on average than the top-performing model.
The security numbers come from Google’s own systems. In Project Fortify, Google’s autonomous scanning, triage, and remediation pipeline, Carbon hit 100% recall on scanning and 95% precision on triage on Google’s internal benchmark, with over 25% relative improvement on Argon’s recall rate when testing open-source repositories. Through Wiz’s Scan for Good program, the report says Carbon surfaced 0-day vulnerabilities in two major SaaS providers in under two hours — compressing weeks of research into an afternoon. Two weeks before general availability, Carbon had already gone to cybersecurity defenders through the Fairwind Program, where Google says it leads benchmarks covering finding (SEC-bench Pro), exploiting (ExploitGym), and patching (CWE-bench v1) vulnerabilities.
The pattern to notice: the strongest cost-per-task evidence sits on benchmarks the vendors themselves designed or co-designed, and the security results sit on Google’s internal benchmark. Those numbers are a reason to run a test, not a reason to switch.
Where the depth actually paid off
The scientific examples show what the extra tokens buy. Google’s researchers used Carbon inside AlphaProof Code to generate millions of lines of Lean proof code, helping formally verify key security properties of Signal’s cryptographic protocol in weeks rather than the years expert teams historically needed. On seven low-homolog protein families from the CASP17 blind test set, Carbon raised retrieved homologs by 40-fold on average, feeding structure prediction without retraining. And Google is expanding its program giving 10,000 scientists free frontier-model access in Antigravity to include Carbon.
Our take: Act now if you run long-horizon agentic work — multi-step debugging, security triage, document-heavy analysis — and stay on Argon if you don’t. The case for acting: access is open today, the intro price equals Argon’s, and a bake-off on a handful of real tasks costs an afternoon and reverses in one config change. Google’s advice that Argon suits most work is credible here because it aligns with Carbon’s mechanism — a model that burns more tokens only wins when the task needs the depth. Assume the 2X-lower-cost claims do not transfer; they were measured by Cursor, Perplexity, Zapier, and Artificial Analysis on their benchmarks, not on buyer workloads. Our threshold: keep Carbon only where it finishes your reference tasks at lower total cost than Argon; if the extra verification loops push it above Argon per finished task, Argon stays. Published per-task costs on real buyer workloads — or the intro price ending before you test — would change that call.
The source page goes further than this article, and it states no rate limits, data-retention terms, or what follows the introductory pricing — the terms that decide whether a small team can adopt it at all.
Sources
View the original post on X
