For two weeks, Gemini 4 Carbon was the model you could read about but not use: Google released it on 28 October 2026 only to a vetted set of cybersecurity defenders in its Fairwind Program. That changed on 10 November 2026, when Google made Carbon — which it calls its strongest model — generally available to developers, enterprises, scientists and consumers, in a single announcement by Koray Kavukcuoglu, Google DeepMind's SVP and chief AI architect.
Carbon is Google's second Gemini 4 frontier model, after Gemini 4 Argon. The division of labour between them is explicit in Google's own announcement: 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" — complex multi-step coding and debugging, open-ended scientific research, and long-horizon agentic workflows. The trade-off is also stated plainly: Carbon uses more tokens than Argon on the same work, spending them on deeper exploration and extended verification loops. Keep that sentence in mind when you read the pricing below.
Who can actually use it now
- Developers: the Gemini API paid tier in Google AI Studio and Android Studio (API model ID
gemini-4-carbon). There is no free-tier access. - Enterprises: Gemini for business, via the API or the model picker, plus Google Antigravity and CodeMender.
- Consumers: Google AI Ultra subscribers only, rolling out in the Gemini app, AI Mode in Search and Antigravity. Pro and free users do not get Carbon, and Google gives no date for wider access.
- Scientists: researchers can apply for six months of free access to both Argon and Carbon at Antigravity's highest usage tier — the program Google opened four weeks ago for 10,000 scientists, now extended to Carbon, with rolling, eligibility-based onboarding.
- Cybersecurity defenders: the Fairwind Program continues on a rolling basis for the cyber-specialised use of the model.
The price has an expiry date
Carbon launches at the same introductory price as Argon: $2 per million input tokens and $10 per million output tokens. The catch is in Google's footnote: after the introductory period ends on 30 November 2026, the price doubles to $4 input / $20 output from 1 December — a schedule Google's own pricing documentation confirms for the Gemini 4 frontier line.
That makes the next three weeks a genuine evaluation window, and it is worth doing the arithmetic before you wire Carbon into anything. At introductory rates, Carbon's output costs a quarter of what Anthropic charges for Claude Fable 5.2 ($8/$40 per million) — a comparison we covered when Fable 5.2 launched earlier this week. But per-token price is not per-task cost: Google itself warns that Carbon burns more thinking tokens than Argon to explore and verify, and thinking tokens are billed as output. A model that is half the price per token but uses twice the tokens on your workload is not cheaper — it is the same bill with better marketing. The rational move for any team is to benchmark Carbon on its own hardest tasks before 30 November, while the experiment costs half as much, and to budget December onwards at $4/$20.
/swarm: one prompt, a team of agents
The feature that signals where Google thinks this is heading is /swarm, now in preview in Google Antigravity. It turns a single prompt into a coordinated team of agents you can track in real time: Carbon takes the roles that demand the deepest reasoning — architectural design, coordination, auditing — while faster, lighter models work in parallel in isolated workspaces, with adversarial verifier agents testing every change. It is the same multi-agent pattern arriving across the industry this month, and it explains why Google is selling Carbon on long-horizon work rather than single answers: the model is priced and positioned as the senior engineer in a swarm, not as a chatbot upgrade.
What Google claims Carbon can do
All of the following are Google's own claims from its announcement; no independent verification of the headline numbers has been published yet.
- Benchmarks: Google says Carbon ranks #1 on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench "at more than 2X lower cost per task compared to the next best external model," and #2 on the new Multimodal Knowledge Agent (MKA) benchmark — launched with Artificial Analysis to test agents reasoning over dense files — at more than 6X lower average cost per task than the top model.
- Cybersecurity: through Fairwind, Google says Carbon leads on finding (SEC-bench Pro), exploiting (ExploitGym) and patching (CWE-bench v1) vulnerabilities. In Google's internal Project Fortify system it claims 100% recall on scanning and 95% precision on triage; Wiz's Scan for Good team used it to find 0-day vulnerabilities in two major SaaS providers in under two hours.
- Science: Google researchers used Carbon in multi-agent setups for formal verification (AlphaProof Code, applied to Signal's cryptographic protocol), enzyme assay development, and protein research — where it says Carbon increased homologs retrieved across seven low-homolog CASP17 protein families by 40-fold on average.
- Safety: because Carbon's cyber capability is stronger, Google says its safeguards blocked 99% of attack trajectories from its adaptive red-teaming agent, cut the per-attempt breach rate to 0.8% (from 3.1% on Argon), and reached a 0% success rate for attackers on the Gray Swan indirect prompt-injection benchmark — a milestone Google says is a first.
Treat the superlatives the way you would any launch-day claim: the direction (a model specialised for verification-heavy, long-running work) is consistent across every example Google gives, but the rankings are the vendor's selection of benchmarks, published on the vendor's schedule.
Why this launch matters
Carbon's general availability is less about one model than about the shape of the frontier race in November 2026. Google now runs two frontier models side by side — Argon for the everyday majority, Carbon for the hardest, longest jobs — at identical prices, and lets the workload decide. That is a different bet from a single flagship, and it lands in the same week Anthropic shipped its own two-name frontier model and turned Claude into an office suite. The practical deadline is the one Google set itself: 30 November, when the introductory price expires and every team that adopted Carbon at $2/$10 finds out what it really thinks the model is worth.
Sources and classification
Classification: Official Announcement (Google, 10 November 2026). Launch facts, availability and pricing are from Google's own announcement and pricing documentation; benchmark, capability and safety figures are Google's own claims, attributed as such above, with no independent verification published yet. Checked 11 November 2026.
- Google Blog, "Gemini 4 Carbon is now available" (Koray Kavukcuoglu, Google DeepMind, 10 November 2026) — read first-hand: general availability to developers, enterprises, scientists and consumers; Argon-vs-Carbon positioning and the higher token usage; Fairwind backstory; /swarm preview in Antigravity; availability via the API in AI Studio and Gemini for business at the introductory $2/$10, and rollout to AI Ultra subscribers in Antigravity, the Gemini app and AI Mode in Search; the introductory-price footnote (pricing rises to $4/$20 after 30 November 2026); benchmark, cybersecurity, science and safety claims; scientist access program: https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-carbon-now-available/
- Google Cloud, Agent Platform / Gemini pricing documentation — Gemini 4 pricing schedule confirming $2/$10 introductory rates through 30 November 2026 and $4/$20 from 1 December 2026 for the Gemini 4 frontier line: https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing
- Google Gemini API documentation — model ID
gemini-4-carbon; paid tier only, no free-tier access (via Google's developer guide and documentation linked from the announcement).