Gemini 4 Carbon leads benchmarks, but it is not ready for production
Google’s new model leads a major independent benchmark, but restricted access means it should not change a production roadmap yet.
What launched
Google announced Gemini 4 Carbon on October 28, positioning it for long-horizon agentic work, coding and cybersecurity. Artificial Analysis scores it at 59 and ranks it first on its Intelligence Index, while The Aggregate records the model as privately accessible.
The operator consequence
Access is limited to selected cyber defenders. There is no public API identifier or general-availability date, so teams cannot test real latency, quotas, reliability or cost per completed task.
What we would do
We would not alter a production roadmap around a leaderboard. We would prepare workload-specific evaluations now, then run Carbon in shadow against the incumbent before considering migration or live routing.
Gemini 4 Carbon sits at the top of the Artificial Analysis index, but for a team maintaining production software it is not yet a product. Google has published pricing and benchmark results while keeping access restricted, with no public API identifier and no general-availability date. We would not delay a migration, redesign an agent or promise a security workflow around it. We would use the announcement to tighten our evaluation set now: real repository defects, patch acceptance, prompt-injection resistance, tool-failure recovery, total tokens, elapsed time and cost per successfully completed task. When Carbon becomes callable, we would run it in shadow beside the incumbent and compare operational behaviour before routing any production traffic. The model may justify a switch; the announcement does not. Until teams can test quotas, latency, observability, data residency and failure modes under their own workload, Carbon is a signal to prepare, not a roadmap dependency.
Turn insight into a working system.
We build provider-neutral evaluation and shadow-routing pipelines that test Carbon against your current model on real tasks before any migration. That includes task-level cost tracking, permissions and audit trails. It is worth talking to us if your team runs coding or security agents that Carbon could affect.