Google opens Gemini 4 Carbon to all paying developers, adds agent swarms

Google made Gemini 4 Carbon available to paying developers, enterprises and Google AI Ultra subscribers on 10 November, two weeks after it first released the model only to a group of cybersecurity defenders through its Fairwind Program. Alongside the wider rollout, Google is previewing a /swarm feature in its Antigravity development platform that turns a single prompt into a coordinated team of AI agents.

This is an update to Unicorn API’s earlier coverage. When Google announced Carbon on 28 October, access was limited to cyber defenders and developers had no date. A few days later, Carbon topped a refreshed third-party hallucination benchmark while still off Google’s public API. That gap is now closed for paying API customers.

What changed since the cyber-defender preview

At announcement, Google gave no date for general developer access, saying only that developers, businesses and consumers would get the model “as soon as possible”, according to RouterPlus. A third-party tracker recorded no public API, playground or weights on announcement day.

Google’s 10 November post says Carbon is now available “to developers, enterprises, scientists, and consumers.” Google calls it “our strongest model”, with frontier performance in cybersecurity and in long-horizon agentic workflows, meaning tasks where an AI agent works through many steps over an extended period, across knowledge work, software engineering and scientific reasoning.

The Verge frames the release as Google rolling the model out “to the general public” after the defender-only phase, and notes Google’s claim that Carbon can compete with the strongest systems from Anthropic and OpenAI.

Where Gemini 4 Carbon is available and what it costs

Google lists these access routes:

• API: Google AI Studio and Gemini for business, for all paying developers and enterprises.
• Consumer and tools: rolling out to all Google AI Ultra subscribers in Google Antigravity, the Gemini app and AI Mode in Search.

The Verge reports it is not clear when, or if, Carbon will reach lower-tier subscribers or free users. For comparison, a LinkedIn post summarising Google’s documentation for the sibling model Gemini 4 Argon said Argon was not available on the API’s free tier.

Item	Gemini 4 Carbon
Introductory input price	$2 per million tokens
Introductory output price	$10 per million tokens
Cached input	$0.10 per million tokens
Scheduled price from 1 December	$4 / $20 per million tokens
Thinking effort levels	Low, medium, high; defaults to medium
Context window (tracker listing)	1M tokens, up to 1M output tokens

Google describes the $2/$10 rate as an introductory price matching Gemini 4 Argon. Simon Willison, who maintains the open-source llm-gemini plugin and released version 0.36 to support Carbon, writes that the launch price matches GPT-6.1 Sol and Claude Sonnet 5.5, and that it is scheduled to double to $4/$20 on 1 December, the same as Claude Opus 5.5. The Verge says the introductory price is around a fifth of the cost of OpenAI’s GPT-6 Astra and a quarter of Anthropic’s Claude Fable 5.2, which launched the day before. It adds a caveat that matters for budgeting: a lower per-token price does not necessarily mean lower costs, because models can use different numbers of tokens for the same task.

That caveat applies directly here. Google itself says Carbon uses more tokens than Argon “to explore thoroughly, reason deeply, and run extended verification loops.”

What /swarm does in Google Antigravity

Antigravity is Google’s agentic development platform, an IDE where AI agents write and run code. The new /swarm feature, in preview, turns “a single prompt into a coordinated team of agents you can track in real time,” according to Google.

Google describes a division of labour: developers can assign Carbon to roles that demand the deepest reasoning, such as architectural design, coordination and auditing, while pairing it with faster, lighter models across parallel isolated workspaces, plus adversarial verifiers (agents whose job is to try to break each change) that test every change.

Google’s own demonstrations include:

• An interactive Three.js 3D exhibition of the Eiffel Tower’s 1887–1889 construction, built from historical photos and engineering plates from a single prompt.
• A 906K-triangle Three.js reconstruction of Google’s Platform 37 office, built from seven public City Council planning PDFs (206 extracted engineering drawing views) and a single prompt, across 10 overnight loops.
• An open-world 3D browser game of Google’s Mountain View campus, with subagents modelling assets in Blender, building the game in Three.js with WebGL shaders, and adding procedural Web Audio effects and Lyria music.

These are maker-selected showcases, not independent tests.

Carbon versus Argon: when Google says to use which

Google says Carbon “pushes the frontier even further” than Gemini 4 Argon, but does not position it as a blanket replacement. 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”, such as complex multi-step coding and debugging, and open-ended scientific research.

Beyond coding, Google pitches Carbon at legal, finance and enterprise operations work: combining visual understanding, reasoning and retrieval over large sets of dense documents, and following strict schemas to generate spreadsheets, presentations and reports. For web apps, Google says it runs build-and-verify loops, recovers from errors and checks both its code and the rendered interface.

i10X News, writing at the original announcement, reached a similar conclusion from the other direction: Carbon “is not a blanket upgrade over Argon” but a specialised tool for long-horizon work and cyber defence.

Google’s benchmark claims and what independent testers show

What Google claims

In its launch post, Google says Carbon ranks first 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. Google also partnered with Artificial Analysis on a new Multimodal Knowledge Agent (MKA) benchmark for agents reasoning over dense files with charts, blueprints and tables; Google says Carbon ranks second there while costing more than 6x less per task on average than the top model.

In pre-release evaluations Google cites, Carbon matched the best model tested on Hebbia’s E2E Financial Agentic evaluation and ranked first out of 23 frontier model configurations on Balyasny Asset Management’s General Financial Intelligence benchmark.

The cybersecurity figures Google published at announcement, 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on CWE-bench, are also Google’s own. AI Model Timeline notes that, per launch coverage, these were run without guardrails.

What third parties show

Vals AI measured Carbon at 63.13% on Terminal-Bench 4.0 at high reasoning effort, ranking it fourth of 59, while Artificial Analysis measured 65.15 on its own run, according to AI Model Timeline. The Verge says the benchmarks Google cited at announcement suggest Carbon is one of the most powerful models on most metrics, behind Fable 5.2, and warns that companies can and do cherry-pick benchmark results.

i10X News reported that Carbon trails competitors on FrontierSWE 2 and OSWorld-2.1, placed fourth on the WebDev Arena leaderboard, and showed slightly higher hallucination rates than Argon on some factual tests in third-party checks. That last point sits awkwardly with Carbon’s lead on Artificial Analysis’s knowledge and hallucination benchmark covered earlier in this thread; the sources do not reconcile the two.

Early hands-on

Willison published outputs at low, medium and high thinking effort and tested Carbon in Antigravity with a complex Three.js animation prompt he had first tried with Claude Fable 5.2. He reports he had to prompt a second time to get a standalone HTML file without external dependencies. It is an informal test, not a benchmark.

What it means for builders, and the open questions

• Budget for the 1 December change. Usage priced today at $2/$10 is reported to cost $4/$20 from 1 December. The same doubling applies to Argon, per a LinkedIn summary of Google’s pricing page. Any cost comparison done this month has a short shelf life.
• Measure cost per task, not per token. Google says Carbon spends more tokens than Argon by design. The Verge notes token counts differ between models on the same task. Google’s own “2x lower cost per task” claims are the figures to check on real workloads.
• Use caching for agent loops. Cached input is listed at $0.10 per million tokens against $2.00 uncached. RouterPlus describes this as 95% off the input price.
• Route by difficulty. Google’s own guidance, and the /swarm design that pairs Carbon with lighter models, both point to using Carbon for planning, coordination and review rather than every call.

Open questions remain. The Verge notes there is no stated timeline for free or lower-tier users. /swarm is a preview inside Antigravity, and the sources do not say whether equivalent orchestration is exposed through the API. RouterPlus noted that Google’s original announcement gave no API model ID. Independent results on several standard benchmarks were still listed as pending by AI Model Timeline as of 31 October.

What to watch next

• Whether Google extends Carbon below AI Ultra, to Pro or free tiers.
• Real-world bills after prices double on 1 December.
• Independent scores from Artificial Analysis and Vals AI on benchmarks still marked pending, and results on Google’s new MKA benchmark from other labs’ models.
• How /swarm moves from preview, and whether developers report the same coordinated-agent results Google showed in its demos.