Google Gemini 4 Carbon release: agent swarm, $2/$10 pricing
Google has made Gemini 4 Carbon generally available to paying developers and enterprises after two weeks of limited access through the Fairwind Program for cybersecurity defenders. The model is available through the Gemini 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. It is also rolling out to Google AI Ultra subscribers in Antigravity, the Gemini app and AI Mode in Search.
Carbon is aimed at long-horizon agentic work, complex coding and open-ended scientific research. Google says Argon remains the ideal choice for the majority of software engineering and knowledge work because Carbon uses more tokens to explore, reason and run verification loops. Alongside the release, Google is previewing /swarm in Antigravity, a feature that turns one prompt into a coordinated team of agents.
In one line: Google has made Gemini 4 Carbon generally available with agent swarm coordination in Antigravity and $2 per million input token pricing after a limited cybersecurity release.
Where Carbon fits
Google positions Carbon as its strongest model for tasks that test the limits of model intelligence, including multi-step coding and debugging, full stack development and scientific reasoning. Argon stays as the default for most routine engineering. The trade-off is token volume, because Carbon thinks for longer and runs extended verification loops, so the same task can cost more even at the same per-token price.
For web app development, Google says Carbon runs build and verify loops, recovers from errors and checks both code and rendered UI. It also lists Carbon at number one on CursorBench 4.0, Perplexity WANDR and Zapier AutomationBench, at more than 2X lower cost per task compared to the next best external model. On the new Multimodal Knowledge Agent benchmark from Artificial Analysis, Carbon ranks second while being more than 6X lower in average cost per task than the top model. These are Google and partner benchmarks, so treat them as vendor reported numbers.
If you are comparing coding tools, the coding and developer tools directory tracks this space.
Agent swarms in Antigravity
The /swarm preview lets a developer send a single prompt and get a team of agents whose work can be tracked in real time. Carbon takes roles that need deep reasoning, such as architecture design, coordination and auditing. Faster, lighter models run in parallel isolated workspaces and act as adversarial verifiers.
Demos include an open world 3D game where a pelican rides a GBike around Google's Mountain View campus, an interactive reconstruction of the Eiffel Tower's construction and a playable 18th century fortepiano with spatial acoustics. This looks like an orchestration layer rather than a single model doing everything, which is close to the ant colony pattern I wrote about earlier.
The cost and benefit of agent swarms is still being debated. The Decoder reports that OpenAI and Anthropic have sounded cautious, and researchers are skeptical too. The main benefit appears to be speed through parallelisation, but orchestration adds complexity. For small teams, this is a watch item rather than an immediate switch.
Cybersecurity and safety claims
Google says the Fairwind Program version of Carbon leads on SEC-bench Pro, ExploitGym and CWE-bench v1 for finding, exploiting and patching vulnerabilities. In Wiz's Scan for Good program, Carbon assessed two major SaaS providers and found zero day vulnerabilities in under two hours. Google's Cloud Vulnerability Research team used Carbon to build proof of concept exploits for two bugs in a popular open-source project, which helped maintainers reclassify one as critical. Project Fortify hit 100 percent recall on scanning and 95 percent precision on triage internally, and improved open-source recall by over 25 percent relative to Argon.
Google also reports low prompt injection risk. On the Gray Swan IPI benchmark, Carbon achieved a zero percent prompt injection success rate, which Google says is the first time any model has hit that milestone. The Decoder points out the test may be saturated, so this does not mean prompt injections are solved. In simulated cyberattacks, Carbon blocked 99 percent of attack paths and cut the breakthrough rate from 3.1 percent with Argon to 0.8 percent. These safety numbers remain vendor reported.
Pricing and competition
The introductory price is $2 per million input tokens and $10 per million output tokens. Simon Willison notes this will double to $4 and $20 on December 1, matching Opus 5.5. The Verge points out that the introductory price is around a fifth the cost of OpenAI GPT-6 Astra and a quarter the cost of Anthropic Fable 5.2. After the price increase, Carbon should still be cheaper per token than frontier rivals, but total cost depends on how many tokens it uses.
The Decoder reports that in Artificial Analysis testing, Carbon cost $7.76 per solved task at the regular price, cheaper than Flash 3.8 and better performing. Opus 5.5 Max cost $10.04, GPT-6 Astra Max $6.25 and Fable 5.2 Max $26.50. Fable 5.2 High sits about two points ahead of Carbon in the overall index at roughly $9 per task. Google's own CursorBench, WANDR and AutomationBench numbers put Carbon ahead on cost per task, so the picture depends on the workload.
What I would try
For client delivery, I would test Carbon on long debugging sessions and multi-step code changes where Argon gets stuck, then compare the total token spend, not just the per-token rate. For RAG chatbots, Carbon's stronger document synthesis and lower prompt injection figures are worth a controlled evaluation, but the general API is not the Fairwind cyber version, so do not assume the same vulnerability hunting performance. I would watch /swarm closely. If Google keeps the orchestration simple enough, it could compress a week of code review and document synthesis into an afternoon, but I would not switch client pipelines to swarm mode until the pricing and failure modes are clear.
Quick answers
What is Google Gemini 4 Carbon?
Gemini 4 Carbon is Google's flagship model for long-horizon agentic work, complex coding, scientific research and cybersecurity. It is generally available at $2 per million input and $10 per million output tokens until December 1, then $4 and $20.
What is the /swarm feature in Google Antigravity?
/swarm is a preview feature that turns one prompt into a coordinated team of agents in Antigravity. Carbon can take architecture, coordination and auditing roles, while lighter models handle parallel workspaces and adversarial verification.
Is Gemini 4 Carbon good for cybersecurity?
Google says the Fairwind Program version of Carbon leads on SEC-bench Pro, ExploitGym and CWE-bench v1. In tests with Wiz, it found zero day vulnerabilities in two SaaS providers in under two hours, and Project Fortify recall improved over 25 percent on open source code.