Gemini 4 Carbon Now Available to Everyone: No. 1 on CursorBench at Less Than Half the Cost per Task

Usually "the most powerful model" means "the most expensive." Gemini 4 Carbon strikes a different deal: Google has released the family's flagship model at the same price as the smaller Argon — $2 per million input tokens and $10 per million output tokens. Access is open through Google AI Studio and Gemini for business to all paying developers, while Google AI Ultra subscribers get it in Antigravity, the Gemini app and AI Mode in Search.

When to Choose Carbon and When to Choose Argon

Google draws the line itself. Argon remains the choice for most development and knowledge-work tasks. Carbon is for cases where you hit the ceiling of a model's intelligence: multi-step debugging, migrations across several packages at once, finding the root cause of a failure that touches multiple subsystems, and open-ended scientific research. Carbon spends more tokens: it explores more widely, reasons more deeply and runs long self-checking loops.

Here is the picture from independent benchmarks. Carbon is first on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench, and the cost per task is more than half that of the best external model. On the new multimodal benchmark MKA, which Google built together with Artificial Analysis, Carbon is second, but on average more than six times cheaper than the leader. These figures come from Google and its partners, so it is worth running the model on your own tasks before adopting it.

/swarm: A Team of Agents from a Single Prompt

The main product novelty for developers is the preview of the /swarm command in Google Antigravity. One prompt turns into a group of agents whose work you can watch in real time. The scheme is straightforward: Carbon takes the roles that demand the deepest thinking — architecture, coordination, auditing. The executors are faster, lighter models working in parallel isolated workspaces, while adversarial verifiers check every change.

The examples from the announcement are impressive: an interactive Three.js reconstruction of a Google office from seven PDFs of blueprints in ten overnight cycles, a 3D game about the Mountain View campus, and a redesign of the Google Merchandise Store in two prompts. These are demonstrations, not a guarantee of results on your codebase, but they show that the model can build and verify working applications, including the rendered interface.

Cybersecurity and Science

Carbon is strong at finding, exploiting and fixing vulnerabilities. In the Project Fortify system it showed 100% recall in scanning and 95% precision in triage on an internal benchmark, while recall on open repositories rose by more than 25% relative to Argon. The Wiz team used it to find 0-day vulnerabilities at two large SaaS providers in under two hours.

In science, Google cites the formal verification of the Signal protocol's cryptography: AlphaProof Code agents generated millions of lines of Lean and proved key security properties in weeks instead of years. For protein structure prediction, Carbon increased the number of homologs found by 40 times on average across seven difficult families from CASP17.

Safety and Limitations

Since the model is stronger at cyberattacks, Google has reinforced its protections. In simulations, Carbon blocked 99% of the attack trajectories of an adaptive red-team agent, and the probability of a breach per attempt fell to 0.8% from 3.1% for Argon. An external review by Trajectory Labs found no new universal jailbreaks over 132 hours. In an agent honesty test of 375 tasks, Carbon passed 92% versus 70% for Argon.

There are limitations too. Resistance to prompt injection, by Google's own admission, remains an area for growth: automated red teaming showed the model can be improved against sophisticated attacks. Full access to cyber capabilities goes through the Fairwind program rather than being open to everyone. The free-access science program is intended for 10,000 scientists.

What to Do in Practice

Don't move everything to Carbon. Keep Argon on routine code and text, and give Carbon the tasks that are expensive in engineers' time: complex migrations, bug investigations, architecture audits, and analysis of large document sets with tables and diagrams. Compare the cost of a solved task, not the price per token: Carbon uses more tokens but may complete the task on the first attempt. Those building agent systems should try /swarm and the "strong model plans and audits, lightweight models execute" scheme.