The Same Model Scored 30 Percent Alone and 100 Percent Inside Nvidia's Agent System
Nvidia's AVO cleared all 183 levels of ARC-AGI-3 using Claude Opus 5, a model that scores about 30 percent on its own. The interesting part is not the score, it is what changed.
Nvidia published results on 21 August showing its research agent AVO clearing every level of the public ARC-AGI-3 benchmark: 183 levels across 25 game environments, with a perfect score of 100.00 on the benchmark’s efficiency metric. The model doing the thinking was Claude Opus 5, which ARC Prize separately measures at roughly 30 percent when it plays alone.
ARC-AGI-3 is an unusual test. An agent is dropped into a small game world with no instructions, no stated rules and no goal. It has to poke at the environment, work out what its actions do, guess what winning looks like, and then do it efficiently. The scoring metric, called RHAE, does not just check whether the agent finished a level; it compares how many moves it needed against a first-time human player. So padding your way to victory with thousands of random actions does not help. AVO solved the whole public set in 6,624 environment actions. VISTA, a rival system running the same model, needed 7,542, about 12 percent more.
AVO stands for Agentic Variation Operators, and it did not start out playing games. Nvidia built it to optimise GPU kernels, the small performance-critical pieces of code that make graphics cards fast. In that job it ran unattended for seven days, tried more than 500 optimisation ideas, committed 40 versions and ended up beating FlashAttention-4 by as much as 10.5 percent. The team then pointed the same agent at ARC-AGI-3 without teaching it anything about games.
The word to know here is harness. A language model on its own answers one question at a time and forgets everything afterwards. A harness is the scaffolding around it: what context it sees, which tools it can call, what it remembers between steps, and what happens when it gets stuck. AVO adds two things in particular. Persistent memory carries findings forward so the agent does not rediscover the same fact ten times. A supervisor watches the overall trajectory and nudges the agent onto a different strategy when it starts going in circles. Nvidia is careful to say this is not a controlled experiment, since the two systems differ in many ways at once, and that these are public-set results, not the private competition set. The honest reading is not “AVO adds 70 points” but “a benchmark number for a model tells you much less than you thought about what an agent built on it can do”.
What this means for you: if you are choosing an AI coding tool or agent, the leaderboard position of the underlying model is only half the story. Two products can run the identical model and behave completely differently, because memory, tool access and error recovery live in the product, not the model. That also explains a common experience: the same model that feels brilliant in one app feels forgetful and scattered in another. For most people nothing changes today, but it is a useful lens for reading the next round of agent announcements.
Sources
Nvidia Paid 6 Billion Dollars for Poolside's Software and 109 of Its People, and Calls It a Licence
No acquisition, no acquihire. Nvidia licenses the platform, invests another billion, and hires most of the team, while Poolside stays independent on paper.
