Nvidia published results on August 21, 2026, for Agentic Variation Operators (AVO), a research system designed to keep an agent productive during long tasks. With Claude Opus 5 underneath, AVO completed all 183 levels across 25 environments in the public set of ARC-AGI-3, an interactive benchmark whose games arrive without instructions.
The 100.00 RHAE score belongs to the full AVO system. ARC Prize separately reported about 30% for Claude Opus 5 at High reasoning effort under a different evaluation setup. The tested configurations differ in tools, memory and task interfaces, so this experiment cannot identify one responsible harness feature. Nvidia did not test the benchmark's private environments.
AVO keeps a persistent record of prior attempts, test results and useful discoveries. A separate supervising agent watches for stalled progress and can redirect the main agent when it repeats unproductive approaches.
Nvidia also tested AVO on GPU-kernel optimization. Over seven days it explored more than 500 possible changes and committed 40 kernel versions. In the tested configurations, the resulting kernels were up to 3.5% faster than cuDNN and 10.5% faster than FlashAttention-4.
AVO is a research system, not a newly released Nvidia product. Its public-benchmark results do not establish security or reliability on every real-world task.