NVIDIA announced on August 21, 2026, that its agent foundation for long-term tasks, "Agentic Variation Operators (AVO)," achieved 100.00 RHAE on the ARC-AGI-3 public set. The foundation model used was Claude Opus 5.
Product LaunchesNvidiaAVOAgentic Variation Operators
NVIDIA's AVO achieves 100% on ARC-AGI-3 public set
This article is a translation. Read the Japanese original
ARC-AGI-3 is an interactive benchmark that evaluates exploration, memory, planning, and action efficiency in game-like environments without explicit instructions or goals. AVO completed the 25 environments and all 183 levels in the public set using 6,624 environment operations.
NVIDIA explains this result not as the performance of a single model, but as the achievement of the entire agent system, which combines persistent memory, tools, feedback, and recovery from failure. For comparison, VISTA also completed all 183 levels using the same Claude Opus 5, but required 7,542 environment operations. However, since the two differ in observation formats, memory, and execution methods, this is not a controlled comparison to measure the contribution of AVO alone.
These results are limited to the ARC-AGI-3 public set and do not reflect performance on the semi-private or private competitive sets.
Source: NVIDIA AVO Reaches 100% on ARC-AGI-3（NVIDIA、2026-08-21）