Gemini 4 Carbon is public: test it as an escalation model, not a default
Google’s new frontier model is available in the paid Gemini API. The launch is worth a pilot, not an immediate default-model migration.
What shipped
Google opened Gemini 4 Carbon to paid API users and business channels on November 10. The model accepts text, images, audio and video with a combined input-output context of up to one million tokens, targeting long-horizon agent workflows, complex coding, research and cybersecurity.
The price is temporary
Standard introductory pricing is $2 per million input tokens and $10 per million output tokens through November 30. Both rates double on December 1, and output billing includes thinking tokens. Carbon is not available on the free API tier.
How we would deploy it
We would not replace an existing model in one move. Carbon should first handle tasks where a cheaper model fails or the cost of an error is high, while shadow evaluations measure completion quality, total token use, latency and reliability. End-to-end task results, not a benchmark table, should determine where it enters production.
Gemini 4 Carbon is worth testing now, but making it the default model would be an operator error. Google has priced it like Argon during the launch period, yet Carbon is explicitly designed to think longer, and all thinking tokens are billed; list-price parity therefore does not equal cost-per-completed-task parity. The benchmark signal is strong enough to justify real work: Carbon leads CursorBench 4.0 in the tested High and Medium configurations and performs strongly in business-agent evaluations. But Google’s own model card also warns of occasional slowness, timeouts and higher token use at stronger thinking levels, while early hands-on testing raises a more important production question: does it stop at the right point without supervision? We would put Carbon behind an escalation policy for difficult migrations, multi-step debugging and audit-heavy research, shadow it against the current model on real production traces, and promote it only where completion quality outweighs latency and token spend. We would not rewrite an agent stack around it before the December 1 price increase has been modelled against actual workloads.
Turn insight into a working system.
We build shadow-routing and evaluation pipelines that test Gemini 4 Carbon on your real tasks before any default-model switch. We add provider fallback, token budgets, permissions and audit trails around the escalation path. Talk to Virtual Arc if your team runs long-horizon agents or complex AI workflows.