September 30, 2026
Cognition boosts AI agent throughput: Vera Rubin delivers up to a 4.8x gain
As of Sep 30, Vera Rubin NVL72 is available on CoreWeave to a limited group of customers, starting with Cognition, the creator of Devin.
Cognition achieved up to 4.8 times the throughput for SWE-2 at the same response speed in tests on Vera Rubin.
More work at once. Cognition compared Vera Rubin NVL72 with GB200 NVL72. For reinforcement learning training of SWE-2, the new system delivered 3.8 times as many output tokens per GPU while maintaining the same level of interactivity. The first result concerns inference, the second training.
CoreWeave is accepting applications for ARENA Pass: 14 days of free access to test your workload alongside the company's engineers. Participants are selected through an application process. To run production workloads, CoreWeave offers a separate discussion of available capacity and launch timing.
From failure to fix. CoreWeave also released Agent Lens in public preview. Claude Code connects to it through Skills and MCP: the AI agent receives a group of similar production failures, examines execution traces, proposes a fix and checks it against a set of test cases. In this cycle, external validation determines whether the work is ready or needs another pass. Insights and evaluations by judge models are free through the end of 2026.
CoreWeave Forge is available as of Sep 30. Weights & Biases users sign in with their existing credentials, and projects and wandb.ai links are preserved. Forge Free costs $0 per month. Pro, for organizations with fewer than 50 employees, starts at $60 per month with a free 30-day trial.
CoreWeave is preparing a separate launch of NVIDIA Vera CPUs: in the company's internal tests, they started isolated environments for agents more than three times faster than an alternative x86 CPU.
