On a Tuesday in November, Google widened the circle around its newest model, Gemini 4 Carbon, carrying it out of a restricted group of testers and into the hands of developers, companies, scientists and everyday users. It is built for the kind of work that unfolds over many steps rather than a single question: writing and debugging software, digging through financial material, and hunting security weaknesses. The company presents it as the stronger option above its sibling Argon, with claims of leading performance that remain, for now, its own to vouch for. Whether those claims hold up under o
Google Opens Gemini 4 Carbon to Developers and Enterprises for Multi-Step Reasoning Tasks
Carbon is for the harder jobs; Argon still covers most of the work
Mark
So what actually changed on Tuesday?
Mimi
Carbon went from a limited release for selected cybersecurity defenders to wider access. Developers, enterprises, scientists and consumers can now reach it, depending on the product.
Mark
How does it differ from Argon, which costs the same?
Mimi
Google says Argon still handles most software and knowledge work. Carbon is for harder jobs: complex coding, debugging, open-ended research that needs repeated verification.
Mark
And the /swarm feature?
Mimi
It's a preview in Google Antigravity. Multiple agents work on parts of a task at once, with Carbon doing system design, coordination and auditing.
Luke
Before we get excited, where do the benchmark claims come from?
Mimi
Google. The report cites pre-release evaluations and benchmarks from Cursor, Perplexity, Zapier, Hebbia and Balyasny.
Luke
So first place on those is Google's statement, not an independent ranking. Nothing here shows who ran the tests.
Mark
What about the cybersecurity numbers?
Mimi
100 per cent recall in vulnerability scanning and 95 per cent precision in triage, for Project Fortify.
Luke
Again, according to the company. We don't know the test set or how hard it was, so those figures can't be read as general performance.
Mark
Is the cost claim solid?
Luke
It's stated as less than half the cost per task of the next-best external model, on specific benchmarks. The report doesn't say which model that was.
Der Puls
- Google has moved Gemini 4 Carbon out of restricted testing and into the hands of developers, enterprises, scientists and consumers at once.
- The model is pitched at long, multi-step jobs, where it must keep reasoning and checking its own work across a chain of actions.
- Google's headline results, including first place on several benchmarks and a 100 per cent recall claim in cybersecurity scanning, come from its own pre-release tests, leaving independent verification thin.
- The new /swarm feature in Google Antigravity lets several AI agents split up a task, with Carbon coordinating design and auditing, and access is spreading to AI Ultra subscribers.
- The next signal to watch is whether outside benchmarks match Google's claims and whether developers really adopt /swarm as the rollout widens.
On a Tuesday in November, Google widened the circle around its newest model, Gemini 4 Carbon, carrying it out of a restricted group of testers and into the hands of developers, companies, scientists and everyday users. It is built for the kind of work that unfolds over many steps rather than a single question: writing and debugging software, digging through financial material, and hunting security weaknesses. The company presents it as the stronger option above its sibling Argon, with claims of leading performance that remain, for now, its own to vouch for. Whether those claims hold up under outside scrutiny is the question that will decide how the moment is remembered.
On Tuesday, Google DeepMind announced that its Gemini 4 Carbon model was leaving its restricted circle of testers and reaching developers, businesses, scientists and consumers. The release is framed around work that cannot be finished in one prompt: software development, financial research and cybersecurity tasks that demand many steps and sustained reasoning.
Google calls Carbon its strongest model yet, pointing to strength in cybersecurity and in long-running agentic workflows, where an AI system works through a long chain of actions. Paying developers can reach it through the Gemini API in Google AI Studio and through Gemini for business, at the same introductory price as Argon: $2 per million input tokens and $10 per million output. Argon remains the choice for most everyday engineering work, while Carbon is meant for the most demanding jobs.
The company supports its launch with benchmark wins, including first place on several coding and automation tests and strong results in financial research. In cybersecurity, its Project Fortify system reported near-perfect recall and high triage precision. All of these figures come from Google's own testing, and independent checks have yet to arrive.
For complex projects, Google is previewing /swarm in Google Antigravity, which lets several agents tackle different parts of a task simultaneously. Access is also reaching AI Ultra subscribers. What happens next will depend on whether outside evaluations confirm the numbers and whether developers embrace the new approach.
On Tuesday, Google DeepMind published a blog post announcing that its Gemini 4 Carbon model was moving beyond a restricted circle of testers and into the hands of developers, businesses, scientists and consumers. The ANI report, carried by The Economic Times from New Delhi on November 11, frames the release around work that cannot be finished in a single prompt: software development, financial research and cybersecurity tasks that take many steps and sustained reasoning.
Google's own description is blunt about where it sees Carbon's strength. The post called it the company's strongest model, with frontier performance in cybersecurity and in long-horizon agentic workflows, meaning jobs where an AI system works through a long chain of actions. Google named knowledge work, software engineering and scientific reasoning as the areas where that applies.
Carbon is built for tasks that need deeper reasoning and repeated checking of its own work, such as complex coding, debugging and open-ended research. Google also said it can work alongside other AI agents, so that separate systems handle separate parts of a larger job.
For paying developers and enterprises, the model is available through the Gemini API in Google AI Studio and through Gemini for business. The introductory price matches that of Gemini 4 Argon: USD 2 per million input tokens and USD 10 per million output tokens. Google positions Argon as still suitable for most software engineering and knowledge work, and Carbon as the option for the more demanding end.
Google backed the launch with benchmark claims. It said Carbon ranked first on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench, at a cost per task that was less than half that of the next-best external model. In financial research, pre-release evaluations showed it matching the best model tested on Hebbia's benchmark and placing first among 23 model configurations on Balyasny Asset Management's financial intelligence benchmark.
In cybersecurity, Google said its Project Fortify system reached 100 per cent recall in vulnerability scanning and 95 per cent precision in triage. Those figures are the company's own. The wider rollout follows an earlier limited release of Carbon to selected cybersecurity defenders through Google's Fairwind programme.
For developers on complex projects, Google is previewing a feature called /swarm in Google Antigravity. It lets multiple AI agents work on different parts of a task at once, with Carbon handling roles such as system design, coordination and auditing. Separately, Google said access is being rolled out to AI Ultra subscribers through Google Antigravity, the Gemini app and AI Mode in Search.
The performance results come from Google's pre-release tests and benchmarks it chose to highlight, so independent verification is still limited. What to watch next is whether outside benchmark results line up with these claims, and how much developers actually use /swarm as the rollout widens.
Bemerkenswerte Zitate
Carbon is Google's strongest model, with frontier performance in cybersecurity and long-horizon agentic workflows across knowledge work, software engineering and scientific reasoning.— Google DeepMind blog post (paraphrased)
Möchten Sie die ganze Geschichte?
Das Original lesen bei The Economic Times ↗