In a season when artificial intelligence is reshaping how work gets done, Google has moved its Gemini 4 Carbon model out of a narrow circle of testers and into the hands of developers, businesses, scientists and everyday users. The company presents Carbon as its most capable system yet, built for the long, patient labours that cannot be finished in one stroke: tangled software, deep research, and the watchful defence of digital systems. Whether that promise holds will be decided less by the company's own confident claims than by the slower verdict of independent testing and by the ordinary wor
Google Opens Gemini 4 Carbon to Developers and Enterprises for Multi-Step AI Work
Carbon is built for work that cannot be finished in a single step
Mark
Google says Carbon is its strongest model. What is actually new here?
Mimi
The reach, mostly. It had gone only to selected cybersecurity defenders through the Fairwind programme. Now paying developers, enterprises, scientists and consumers can use it.
Mark
And what is it meant to do that Argon doesn't?
Mimi
Harder, multi-step work: complex coding, debugging, open-ended research that needs repeated verification. Google says Argon still suits most software and knowledge work.
Luke
Before we go further, how much of the performance story is independently confirmed?
Mimi
The report attributes the rankings to Google. First place on CursorBench 4.0, WANDR and AutomationBench is the company's claim.
Luke
Same for the cybersecurity numbers. 100 per cent recall and 95 per cent precision come from Google about its own Project Fortify system. The report doesn't say how the testing was set up.
Mark
The pricing is the same as Argon, though. That seems notable.
Mimi
It is. $2 per million input tokens and $10 per million output. Google also says Carbon costs less than half as much per task as the next-best external model on those benchmarks.
Luke
That cost comparison is Google's too, and it's limited to the benchmarks named. Also, the finance results were pre-release evaluations, so they predate this rollout.
Mark
What about '/swarm'?
Mimi
It's a preview in Antigravity, with several agents working at once and Carbon handling design, coordination and auditing. How well it works in practice isn't known yet.
Der Puls
- Google has opened its strongest model, Gemini 4 Carbon, to paying developers and enterprises, shifting it from guarded trials into open commercial use.
- Its benchmark victories are mostly self-reported, leaving rival labs and independent testers to decide whether the claimed leads truly hold.
- Carbon is pitched at multi-step coding, financial research and cybersecurity, where tasks unfold over many stages rather than a single reply.
- Priced the same as the lighter Gemini 4 Argon at $2 and $10 per million tokens, Carbon is being steered toward the hardest jobs while Argon handles everyday work.
- Attention now turns to the '/swarm' preview in Google Antigravity, where several AI agents will divide large projects, and to whether it performs on real codebases.
In a season when artificial intelligence is reshaping how work gets done, Google has moved its Gemini 4 Carbon model out of a narrow circle of testers and into the hands of developers, businesses, scientists and everyday users. The company presents Carbon as its most capable system yet, built for the long, patient labours that cannot be finished in one stroke: tangled software, deep research, and the watchful defence of digital systems. Whether that promise holds will be decided less by the company's own confident claims than by the slower verdict of independent testing and by the ordinary work of people who put the model to use.
On Tuesday, Google DeepMind announced that its Gemini 4 Carbon model was leaving its small group of testers and reaching developers, businesses, scientists and consumers. Google calls Carbon its strongest model, one designed for long-horizon work in cybersecurity, software engineering and scientific reasoning.
Carbon is meant for tasks that cannot be completed in a single step, such as complex coding, debugging and open-ended research, where the model must reason deeply and check its own work again and again. It can also work alongside other AI agents, each taking a piece of a larger job.
Access runs through the Gemini API in Google AI Studio and through Gemini for business, priced at the same rate as the lighter Gemini 4 Argon. Google keeps Argon for most everyday engineering and knowledge work and reserves Carbon for the most demanding cases.
The company's benchmark claims are largely its own. Google says Carbon leads several coding and automation tests at under half the cost per task of the next-best rival, and it reports strong results in financial research and a vulnerability-scanning system called Project Fortify.
For large projects, Google is previewing '/swarm' in Google Antigravity, where Carbon coordinates, designs and audits while multiple agents work at once. What remains open is whether independent testing confirms the rankings, and how the feature fares once real developers rely on it.
On Tuesday, Google DeepMind published a blog post announcing that its Gemini 4 Carbon model was moving beyond a small circle of testers and into the hands of developers, businesses, scientists and consumers. The company called Carbon its strongest model, one with what it described as frontier performance in cybersecurity and in long-horizon agentic workflows across complex knowledge work, software engineering and scientific reasoning.
The model is built for jobs that cannot be finished in a single step: complex coding, debugging and open-ended scientific research, where the work calls for deeper reasoning and repeated verification. Google also said Carbon can operate alongside other AI agents, so that different systems take on separate parts of a larger task.
Access comes through the Gemini API in Google AI Studio and through Gemini for business, for paying developers and enterprises. The introductory price matches that of Gemini 4 Argon: $2 per million input tokens and $10 per million output tokens. Google positions Argon as the model that remains suitable for most software engineering and knowledge work, with Carbon reserved for the more demanding cases.
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 it put at less than half that of the next-best external model. In finance, pre-release evaluations showed Carbon matching the best model tested on Hebbia's financial research 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, as the wire report attributes them to Google.
For developers on large projects, Google is previewing a feature called '/swarm' in Google Antigravity. It lets multiple AI agents work on different parts of a task at the same time, with Carbon handling roles such as system design, coordination and auditing.
The wider release follows an earlier, limited rollout of Carbon to selected cybersecurity defenders through Google's Fairwind programme. Google said access is also being rolled out for AI Ultra subscribers through Google Antigravity, the Gemini app and AI Mode in Search.
What to watch next is whether independent testing bears out the benchmark rankings, which come largely from Google, and how the '/swarm' preview performs once developers start using it on real projects.
Bemerkenswerte Zitate
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.— Google DeepMind blog post
Möchten Sie die ganze Geschichte?
Das Original lesen bei The Tribune ↗