ARC-AGI-3 is mean. The agent gets dropped into a 2D game with no explanation, no rules, no goal. It has to figure out for itself what even counts as a reward and how to make progress. That’s exactly where most models fall apart.
NVIDIA has now shown an agent called AVO that gets a perfect score on the public part of the benchmark: all 183 levels across 25 environments, an RHAE of 100.00. The twist is what’s running inside. The reasoning model in AVO is Claude Opus 5 – the same model that, on its own in a high-reasoning setting, gets to only about 30 percent according to ARC Prize.
The difference isn’t the model
Same model, 30 one way, 100 the other. What sits in between isn’t a better brain, it’s a better setup. NVIDIA calls it the harness, and it has three parts: a memory that carries trials and results across runs; a supervisory layer that spots dead ends and corrects course; and tools the agent uses to check its own actions.
To be fair, NVIDIA says so itself: this isn’t a clean controlled comparison, and you can’t precisely quantify AVO’s contribution alone. The predecessor VISTA already hit 100 with the same Claude Opus 5; AVO just needs fewer steps. Still, the message is clear: before you buy the next, more expensive model, rebuild the scaffolding around the one you have.
Why this makes me happy
I’ve been saying the same thing for a while: my experience plus Claude Code, together, are unbeatable. ARC-AGI-3 is basically the nerdy proof. The model is the raw thinking power. But whether that turns into something useful is decided by the system around it – memory, oversight, self-checking, clear tasks.
For you, that means: when an agent gives you bad results, the model isn’t automatically to blame. Often the scaffolding is missing. A bit of structure, a second look by the agent at its own work, a memory that spans multiple steps – in practice that often beats jumping to the next model tier. And it usually costs thinking, not money.
Sources: