Google has thrown open the doors to Gemini 4 Carbon, its most capable model to date, making it available to developers, enterprises, scientists and consumers just two weeks after a limited rollout to a hand-picked group of cybersecurity defenders through the Fairwind Program. The release arrives alongside /swarm, a preview feature in Google Antigravity that converts a single prompt into a coordinated team of agents whose work can be tracked in real time.
- Broad availability: Carbon is live via the API in Google AI Studio and Gemini for business for all paying developers and enterprises, and is rolling out to Google AI Ultra subscribers across Antigravity, the Gemini app and AI Mode in Search.
- Introductory pricing: The model launches at the same rate as Gemini 4 Argon — $2 per million input tokens and $10 per million output tokens — with those prices set to double on December 1, 2026.
- Agent orchestration: The /swarm preview lets developers assign Carbon to deep-reasoning roles such as architectural design, coordination and auditing, while lighter models run in parallel isolated workspaces alongside adversarial verifiers.
A specialist model, not a daily driver
Google is explicit that Carbon is not meant to replace its generalist sibling. According to the company, Gemini 4 Argon remains the ideal choice for the majority of software engineering and knowledge work, while Carbon targets tasks that push a model’s intelligence to its limits: complex multi-step coding and debugging, open-ended scientific research and discovery. To get there, Carbon burns through more tokens than Argon, exploring thoroughly, reasoning deeply and running extended verification loops.
The division of labour mirrors the strategy Google sketched when it launched Argon as its mainstream model, and it was already visible when Carbon first surfaced with restricted cybersecurity access. In that earlier phase, Google said Carbon led benchmarks for finding, exploiting and patching vulnerabilities. The company now reports that in Project Fortify, its autonomous scanning, triage and remediation system, Carbon hit 100% recall on scanning and 95% precision on triage on an internal benchmark, with a relative improvement of over 25% on open source code repository testing.
Benchmarks and the economics of intelligence
Google claims 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. It also partnered with Artificial Analysis on a new Multimodal Knowledge Agent benchmark for dense documents, charts and schematics, where Carbon places second while costing more than six times less per task than the top performer. In pre-release evaluations, Google says Carbon matched the best model tested on Hebbia’s E2E Financial Agentic evaluation and topped 23 frontier model configurations on Balyasny Asset Management’s General Financial Intelligence benchmark.
Independent documentation adds practical detail. Carbon supports three effort levels — high, medium as the default, and low — and a context window of roughly one million tokens, with up to 65,536 output tokens per synchronous request. Thinking tokens bill as output, implicit caching kicks in from 4,096 prompt tokens, and batch and flex tiers run at $1 input and $5 output. Regional endpoints on the Gemini Enterprise Agent Platform carry a 10% premium over global pricing.
Carbon is our strongest model with frontier performance in cybersecurity capabilities and long-horizon agentic workflows across complex knowledge work, software engineering, and scientific reasoning.
Swarm logic and what comes next
The /swarm feature extends an orchestration pattern Google has already pushed into the smart home, now aimed squarely at software development. Google’s showcase examples are deliberately theatrical: teams of Carbon agents that reconstructed the Eiffel Tower in 3D from historical photographs, or generated an open-world video game from scratch. Beyond the demos, the architecture matters — pairing a heavyweight reasoner with fast, cheap workers and adversarial checkers is a template for how multi-agent systems will be built and billed.
Google is also courting the research community with six months of free access for 10,000 scientists, a move that could seed Carbon deep inside academic workflows before competitors respond. The strategic read is that the real story here is not raw capability but price and cadence. A flagship model priced identically to its generalist sibling, and pitched as several times cheaper per task than its nearest rival, signals that the AI frontier is now contested on cost-effectiveness as much as on intelligence. Carbon’s reign may be brief — Anthropic has already answered with Claude Fable 5.2 and Mythos 5.2 — but the pressure it exerts on pricing benefits users directly, and that race has only just begun.