Google presses its advantage in the model race
Google has opened its most capable model yet to a wide audience. Gemini 4 Carbon is now available to paying Gemini API developers, enterprise customers and subscribers to Google AI Ultra, and it is rolling out across the Gemini app, AI Mode in Search and Google’s Antigravity coding platform.
The company describes Carbon as a model for long-horizon work: tasks that need many steps, sustained reasoning and a lot of checking along the way. Examples include complex coding and debugging and open-ended scientific research. Carbon follows the earlier Gemini 4 Argon model, and the release arrives after a two-week preview that was limited to cybersecurity professionals.
One prompt, many agents
The most eye-catching addition is /swarm, a feature in preview inside Antigravity. It takes a single instruction and turns it into a coordinated team of agents that users can watch in real time.
Google’s approach splits the work by difficulty. Carbon takes the roles that need the deepest reasoning, such as designing the architecture, coordinating the effort and auditing the results. Faster, lighter models then work in parallel in isolated workspaces, while separate verifier agents try to break each change before it is accepted. Google says it tested /swarm internally for five weeks before opening the preview.
In demonstrations, the company showed Carbon modelling 3D assets in Blender and building browser-based scenes. It also worked with the analytics firm Artificial Analysis on a new benchmark designed to test how well agents reason across dense files that contain charts, images, blueprints and tables. GitHub has said the model is aimed at complex coding and agent tasks, with a focus on secure code generation.
A price designed to turn heads
Pricing may matter as much as performance. Carbon is launching at $2 per million input tokens and $10 per million output tokens, the same introductory rate as Argon. That is a fraction of the cost of several rival flagship models. The Verge reported that it is around a fifth of the price of OpenAI’s GPT-6 Astra.
There is a catch. The introductory rate runs only until November 30. After that, Carbon will cost $4 per million input tokens and $20 per million output tokens. Even at the higher price, reports comparing published rates suggest it will remain cheaper than many frontier competitors on a per-token basis.
Per-token cost is not the whole story, though. Different models use different numbers of tokens to finish the same job, so the real cost of a task can vary. A cheaper model that needs more attempts or longer outputs may not save money in practice. Developers will need to test it on their own workloads.
Where Carbon sits among its rivals
Google’s own benchmark claims place Carbon among the most powerful models on the market. According to The Verge, it still trails at least one rival flagship on most measures. Google says it ranks first on several coding and automation tests at much lower cost per task than the next-best outside model, though those are the company’s figures and independent testing will take time.
The release does show how competitive the field has become. Google, OpenAI, Anthropic and others are shipping new models at a rapid pace, and each is competing for the developers and companies that are building software on top of them. Coding agents are the prize that stands out: they are the area where businesses can see the clearest near-term return on AI spending.
The infrastructure question
Cheaper, more capable models arrive at a sensitive moment for the industry. Investors have spent recent weeks debating how long the AI building boom can last, and whether the huge sums going into chips, memory and data centres will pay off. Lower prices from model makers could expand demand by making AI affordable to more companies. They could also squeeze margins for those that sell access to it.
For developers, the immediate decision is practical. Many will try /swarm and the new pricing before the introductory window closes. Others will wait to see how reliably multi-agent systems perform on real projects, where mistakes can multiply as quickly as the agents do.
What to watch
- Independent benchmarks. Third-party tests will show whether Carbon’s cost advantage holds in practice.
- The end of the promotional price. Usage after November 30 will reveal how price-sensitive customers really are.
- Reliability of agent teams. The /swarm preview will be judged on whether coordinated agents can finish large tasks without constant human fixes.
