Claude Opus 3 hits 100% on ARC-AGI — but only with Nvidia's custom harness
nvidia researchers boosted claude opus 3 from a 30% to a 100% score on the arc-agi benchmark using a custom harness with a supervisor component, proving that scaffolding matters more than model size for long-horizon tasks.
nvidia research revealed that a custom harness called agentic variation operators (avo) allowed claude opus 3 to score 100% on the arc-agi benchmark, a significant jump from the 30% baseline. the system utilized a supervising agent to correct the main model's direction, solving the issue of ai going off-track during long-horizon tasks. this finding challenges the industry's focus on model size, suggesting that the 'harness' or scaffolding managing memory and feedback is the true driver of agentic performance. it signals a shift for developers to prioritize runtime infrastructure over raw model capability.