Claude Opus 5 scored 30% on ARC-AGI-3. Wrapped in Nvidia’s AVO, it hit 100%.
The New Stack Adrian Bridgwater ● Covered by 4 sources
Claude Opus 5 scored 30.2% on ARC-AGI-3 alone. Put inside Nvidia’s AVO agent system, it reached 100%.
Based on reporting by The New Stack, Adrian Bridgwater — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Nvidia says the model is only part of the story. In a blog post from Friday, a five-person team of engineers, machine learning specialists and AI research interns argued that its Agentic Variation Operators system can turn a decent model score into sustained autonomous performance.
The example they chose is blunt. Claude Opus 5 reportedly scored 30.2% on the ARC-AGI-3 public set on its own. Run through AVO, Nvidia says, it reached 100.00 RHAE across all 25 environments and completed all 183 levels. The team’s point is not that the model suddenly became smarter. It is that the surrounding system kept the work moving.
AVO is built for long-running agent work: inspecting and editing code, running commands, reading documentation and checking results by execution. Nvidia says the system matters most when tasks stretch over long horizons and the model has to keep going with only a partial view of the problem. That is where persistent memory and supervision come in. Memory preserves prior work, compiler output, profiler results and accumulated reasoning. Supervision, in this case a programmatic software module, watches the search and can step in when progress stalls.
The same architecture was first shown on software engineering and GPU-kernel optimization tasks, and Nvidia is now using that history to argue for transfer. The company says the pattern is similar even when the surface problem changes: form a hypothesis, act, observe, update state, then keep going. In ARC-AGI-3, the feedback comes from environment transitions and action outcomes rather than compilers and benchmarks, but the loop is supposed to be the same.
Nvidia also paired AVO with GPT-5.6 Sol on a challenging subset of games. In those tests, Sol reached matched levels faster in some cases, while Opus used fewer environment actions in matched-level comparisons. The company says those are preliminary results and not scores from the semi-private or fully private competition sets. The broader claim is simpler: model capability matters, but the agent harness decides how much of that capability survives contact with reality.
My take — AI-written commentary, not fact-checked reporting
This is the part of AI everyone keeps pretending is a side quest: the system around the model. Nvidia is basically saying the same thing the industry keeps learning the hard way — raw model scores are cute, but memory, tools, and supervision are what turn a demo into a worker. The hype machine loves naming the model; the boring plumbing is where the actual leverage lives.
Read more about this at: The New Stack
