Google Opens Gemini 4 Carbon to Paying Developers, With Prices Doubling Dec 1

What happened

On Nov 10, Koray Kavukcuoglu, Google DeepMind's senior vice president and chief AI architect, said in the official blog that Gemini 4 Carbon is now open to developers, enterprises, scientists and consumers. Google calls Carbon its strongest model to date, built for cybersecurity and long-horizon agentic work across complex multi-step coding, debugging and open-ended scientific research.

Access runs through three routes: developers use the paid tier of the Gemini API in Google AI Studio and Android Studio; enterprises connect through Gemini for business; AI Ultra subscribers get it in the Gemini app, AI Mode in Search and Antigravity. Google is also previewing /swarm in Antigravity, which turns a single prompt into a coordinated team of agents that can be tracked in real time, with Carbon assigned to roles that need the deepest reasoning, such as architectural design, coordination and auditing.

The timeline matches Argon's exactly: Argon was announced on Sep 30 and opened widely on Oct 13, a 13-day gap; Carbon was announced on Oct 28 and opened widely on Nov 10, the same 13-day gap. At launch on Oct 28, Carbon went only to a subset of cybersecurity defenders in the Fairwind Program, with no date set for wider availability; the cybersecurity variant stays inside Fairwind, a separate access path from the public model.

The scientist program grows at the same time: four weeks ago Google opened applications to give 10,000 scientists six months of free access to Argon in Antigravity's highest usage tier, and thousands have already been onboarded; from Nov 10 the program covers both Argon and Carbon, with applications continuing on a weekly rolling basis.

Pricing and access

API pricing carries over Argon's introductory rate: $2 per million input tokens and $10 per million output tokens, with the introductory period ending Nov 30 and rates rising to $4 and $20 from Dec 1. Google's pricing page says the model is not available on the free tier, and reasoning tokens are billed as well; cached input tokens cost 95% less.

On the consumer side the bar sits at the subscription tier: Google AI Ultra subscribers can use Carbon in the Gemini app, where the model picker labels it 'Most complex work'; 9to5Google reported that the menu dropped Gemini 3.1 Pro at the same time. Free, AI Plus and AI Pro users do not get it, and Argon remains the highest model on the Pro plan; Google has not said when, or if, Carbon will reach those plans. Ultra starts at $99.99 a month in the US.

For comparison, The Verge put Carbon's introductory price at roughly a fifth of GPT-6 Astra and a quarter of Claude Fable 5.2, which launched a day earlier at $8 per million input tokens and $40 output. At the Dec 1 standard rate, Carbon matches Claude Opus 5.5 at $4 and $20. Google notes in its blog that Carbon uses more tokens than Argon for deeper reasoning, so per-task costs differ even at the same per-token rate.

What people are saying

Google CEO Sundar Pichai said on X on Nov 10 that Carbon is rolling out today, built for sustained, long-horizon work across complex software engineering, cybersecurity and scientific discovery. Google said in its blog that Carbon ranks first on three independent benchmarks — CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench — at less than half the cost per task of the next best external model.

Independent evaluator Artificial Analysis scores Carbon 59 on its Intelligence Index v5.0, first overall, ahead of Claude Opus 5.5 at 58 and GPT-6 Astra at 57; on the same firm's coding-agent index, Carbon inside the Antigravity CLI scores 59, below Claude Code with Opus 5.5 at 62.

Early testers reported measured results: Google said Carbon reached 100% recall on scanning and 95% precision on triage inside Project Fortify, its internal autonomous scanning, triage and remediation system; its Cloud Vulnerability Research team used Carbon to prove two bugs in popular open-source software were exploitable, prompting the maintainers to reclassify them as critical and issue patches; Wiz's Scan for Good program used Carbon to find 0-day vulnerabilities at two large SaaS providers in under two hours.

The safety figures come from the Carbon safety and security report Google published alongside the release: in simulated cyberattack testing, Carbon blocked 99% of attack trajectories, with a per-attempt breach rate of 0.8%, down from 3.1% on Argon; Gray Swan testing put automated cyberattack query success at 2.17%, down from 9.86% on Argon; on a 375-task agentic honesty benchmark it passed 92% of tasks, up from 70% for Argon.

Key facts

1. Availability: Paid API tier (AI Studio and Android Studio), enterprises (Gemini for business), and AI Ultra subscribers (Gemini app, AI Mode in Search, Antigravity); free and AI Pro users do not have access yet
2. Pricing: $2 per million input tokens and $10 output through Nov 30, rising to $4 and $20 from Dec 1; not available on the free tier, and reasoning tokens are billed as well
3. Timeline: Announced Oct 28 for Fairwind Program cybersecurity defenders only; opened to paying developers, enterprises and AI Ultra subscribers on Nov 10, the same 13-day gap as Argon's Sep 30 to Oct 13 rollout
4. Context window: 1,048,576 tokens, shared by input, thinking and output
5. Free for research: Google is giving 10,000 scientists six months of free access in Antigravity's highest usage tier; from Nov 10 the program covers both Argon and Carbon
6. Cyber variant: The defender version stays inside the Fairwind program, a separate access path from the public model
7. Benchmarks (Google, Oct 28): Artificial Analysis Intelligence Index v5.0: 59, first overall; Terminal-Bench 4.0: 65.2%; SEC-bench Pro: 88.2% and CWE-bench v1: 71%; FrontierSWE 2: 61.2%, behind GPT-6 Astra at 65.5% and Claude Opus 5.5 at 62.3%
8. New tool: Google is previewing /swarm in Antigravity, turning one prompt into a coordinated team of agents tracked in real time
9. Safety figures: Simulated attack testing blocked 99% of attack trajectories, per-attempt breach rate 0.8%; agentic honesty benchmark pass rate 92%