# Google's Gemini 4 Carbon adds teams of agents that work together on tasks

> Google's Gemini 4 Carbon introduces teams of agents that can work together on tasks, as major model releases keep accelerating.

- Status: reported (One outlet, not independently verified yet.)
- Topic: Models
- Published: 2026-11-15T20:29:01+00:00
- Updated: 2026-11-15T21:10:00.773694+00:00
- Web page: https://ai-news.ikangai.com/story/google-s-gemini-4-carbon-adds-teams-of-agents-that-work-together-on-tasks/
- Seen in: THE DECODER
- Change: capability (New capability)
- Takes effect: 2026-11-10
- Applies to: Gemini 4 Carbon, /swarm

## What changed

Gemini 4 Carbon is now generally available, and Google previewed /swarm, a feature that turns one prompt into a coordinated team of agents.

Google made Gemini 4 Carbon generally available on November 10, 2026. Google calls it the strongest model in its Gemini 4 family and previewed a feature called /swarm, which turns a single prompt into a coordinated team of agents. The release came about two weeks after Google first unveiled Carbon to a limited group of cybersecurity defenders through its Fairwind Program. Anthropic had released Fable 5.2 by then, and The Decoder noted that it outperforms Carbon on the Artificial Analysis test even at lower reasoning intensity.

Google positions Carbon alongside Gemini 4 Argon, which it released earlier, rather than as a replacement. Google says Argon remains the better choice for most software engineering and knowledge work. Carbon is built for tasks that test the limits of a model's intelligence, such as multi-step coding and debugging and open-ended scientific research. The trade-off shows up in the token count. Carbon uses more tokens than Argon to explore problems, reason deeply and run extended verification loops. Both models carry the same token price, so The Decoder argued that the total bill for comparable work is likely higher on Carbon.

The introductory price for paying developers and enterprises is $2 per million input tokens and $10 per million output tokens. It is available through the Gemini API in Google AI Studio, Android Studio and Gemini for business. The fine print matters here. After the introductory period expires on November 30, 2026, the public-access Carbon model costs $4 per million input tokens and $20 per million output tokens. Google AI Ultra subscribers get access in Google Antigravity, the Gemini app and AI Mode in Search.

The /swarm feature, currently in preview in Google Antigravity, assigns Carbon to the roles that require the deepest reasoning, such as architectural design, coordination and auditing. Faster, lighter models run in parallel isolated workspaces, and adversarial verifiers test every change. Google demonstrated it with a 3D game in which a pelican rides a GBike across a rendering of its Mountain View campus. It also showed an interactive Eiffel Tower exhibition built from historical photos and a playable 18th-century fortepiano. The Decoder reported that the cost-benefit case for agent swarms is being debated across the industry. It also reported that OpenAI and Anthropic have recently sounded cautious, and that the main benefit appears to be speed through parallelization.

Google claims 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. The Decoder reported that Carbon scores just under 60 percent on CursorBench 4.0 at roughly $8 to $10 per task. In the Artificial Analysis test, Carbon cost $7.76 per solved task at standard pricing. That is below Opus 5.5 Max at $10.04 and Fable 5.2 at its maximum tier at $26.50, but above GPT-6 Astra Max at $6.25. Fable 5.2 at its High tier costs about $9 per task and sits two points ahead of Carbon on the overall index, 62 to 60. On Google's new Multimodal Knowledge Agent benchmark, which Google built with Artificial Analysis, Carbon ranks second, behind Fable 5.2.

On cybersecurity, Google says Carbon found zero-day vulnerabilities in two large SaaS providers in under two hours while working with Wiz's Scan for Good program. Google's Cloud Vulnerability Research team used the model to build proof-of-concept exploits for two bugs in a popular open-source project. That work helped maintainers reclassify one of them as critical. In Project Fortify, Google's autonomous scanning, triage and remediation system, Carbon reached 100 percent recall on scanning and 95 percent precision on triage on an internal benchmark. On open-source repository testing, Google reports an over 25 percent relative improvement on Argon's recall rate. The Decoder noted that Google's leads on SEC-bench Pro, ExploitGym and CWE-bench v1 apply only to the version available through the Fairwind Program. It also noted that Google doesn't share specific numbers for them.

Carbon blocked 99 percent of attack trajectories from an adaptive red-teaming agent that revisits its plans after each block. It also cut the per-attempt breach rate to 0.8 percent from 3.1 percent on Argon. In Gray Swan's testing, the success rate of automated cyberattack queries fell to 2.17 percent from 9.86 percent. On a 375-task agentic honesty benchmark, Carbon passed 92 percent of tasks, against 70 percent for Argon. It declined to fill in missing experimental data 98 percent of the time and flagged methodological flaws 68 percent of the time. External testers at Trajectory Labs reported zero new universal jailbreaks across 132 hours and over 13,000 requests.

Google says Carbon achieved a 0 percent prompt injection success rate on the Gray Swan IPI benchmark, the first time any model has reached that milestone. Google's own post adds that prompt injection robustness remains an area for continued improvement. The Decoder said the result doesn't mean prompt injection is solved and that the test may simply be saturated. It noted, though, that some competitors are still more vulnerable.

On science, Google describes a framework called AlphaProof Code. In it, Carbon helped generate millions of lines of Lean code to prove key security properties of Signal's cryptographic protocol in weeks. Google says that task historically takes expert teams years. In protein structure work, Google reports that across seven low-homolog protein families from the CASP17 blind test set, Carbon increased the number of retrieved homologs 40-fold on average. Google's post notes that broader benchmarking of that result is still underway.

## Why it matters

Agent-team features built into a flagship model could change how companies split long, multi-step work between software agents and people.

## What you can do

Test whether multi-agent workflows on Gemini 4 Carbon beat single-agent runs on your own tasks before adopting them.

## Original source

- [Gemini 4 Carbon is now available](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-carbon-now-available/)

## Further reporting

- [Gemini 4 Carbon arrives just in time to feel like yesterday's news in the fastest model cycle yet](https://the-decoder.com/gemini-4-carbon-arrives-just-in-time-to-feel-like-yesterdays-news-in-the-fastest-model-cycle-yet/)

---

AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it.

Cite as: IKANGAI AI News, "Google's Gemini 4 Carbon adds teams of agents that work together on tasks", 2026-11-15, https://ai-news.ikangai.com/story/google-s-gemini-4-carbon-adds-teams-of-agents-that-work-together-on-tasks/
