AI Agents August 24, 2026

Nvidia's AVO Put Claude Opus 5 at 100% on ARC-AGI-3's Public Set

NVIDIA reports that its AVO agent system reached a 100.00 RHAE score across ARC-AGI-3's public demo. The result is a striking system-level outcome, but NVIDIA cautions that it is not a controlled measurement of scaffolding alone.

A benchmark result from NVIDIA has put a sharp new question in front of AI builders: how much of an agent's performance belongs to the model, and how much belongs to the system built around it? On August 21, NVIDIA reported that AVO, its Agentic Variation Operators architecture using Claude Opus 5, reached a 100.00 RHAE score across all 183 levels in ARC-AGI-3's 25 public demo environments.

ARC-AGI-3 is deliberately unlike a static question set. Agents explore unfamiliar interactive environments without instructions, rules, or stated goals, then have to infer what matters and act on it. NVIDIA says AVO completed the public set in 6,624 environment actions, compared with 7,542 actions reported for VISTA. That is a notable public demonstration of an agent system, though the two systems differ in more than one way.

The 30% Comparison Has a Boundary

The headline number invites an easy story: Claude Opus 5 was around 30% on ARC-AGI-3, then a loop made it 100%. The underlying facts are more careful. ARC Prize reports Claude Opus 5 (High) at 30.2% on its ARC-AGI-3 result page. NVIDIA reports 100.00 RHAE for AVO with Claude Opus 5 on the public demo.

NVIDIA explicitly says that this is not a controlled ablation of AVO. The ARC Prize reference used High reasoning effort, while NVIDIA's run used a different reasoning setting and a different agent, observation, memory, context, and evaluation setup. The company says the comparison should not be read as a direct measurement of AVO's contribution. It is a system-level result, not proof that one isolated loop caused a 70-point lift.

That qualification makes the result more useful, not less. It puts the engineering question in the right place: a model's score can depend materially on the tools, state, feedback cycle, and evaluation harness that surround it.

What AVO Adds

NVIDIA describes AVO as a general-purpose coding-agent architecture with persistent memory, supervision, and tool use. Its agent can inspect an environment, develop and test hypotheses, edit code, run commands, consult documentation, and validate the result. Observations and failures from one turn inform the next rather than disappearing after a single response.

That is the scaffolding argument in its strongest defensible form. The model is part of the capability, but so are the loops that give it context, let it take actions, preserve what it learned, and check whether those actions worked. For teams deploying agents in production, those are design choices they control directly.

Why the Public Set Matters — and Why It Isn't the Final Word

AVO's result is on the public demo, not a semi-private or private evaluation. NVIDIA added an editorial clarification to make that distinction explicit. The public environments are available for researchers to inspect and run repeatedly, so a result there can demonstrate a working system but cannot, by itself, establish how well it generalizes to held-out tasks.

ARC Prize's technical report is even blunter: it says public-set results are not used on the official leaderboard because public environments can be directly targeted during system development and are materially easier than the private set. The official 2026 competition evaluates agents through a designated Kaggle submission under its own rules.

So 100% is neither nothing nor a solved general-intelligence claim. It is a reproducible-looking public-system result with a clear boundary around what it establishes.

What to Watch Next

The next useful evidence would be an evaluation on held-out environments, alongside enough implementation detail for others to reproduce the result. Cost, action count, model configuration, and the degree to which a harness is general-purpose all matter when comparing agent systems.

NVIDIA says AVO was built for long-horizon coding work, including GPU kernel optimization, rather than only for ARC-AGI-3. If the same design produces robust results outside this public benchmark, the lesson will be broader: capability is increasingly a property of the full agent system, not a number that lives in model weights alone.

Sources

NVIDIA Developer Blog — NVIDIA AVO Reaches 100% on ARC-AGI-3: https://developer.nvidia.com/blog/nvidia-avo-reaches-100-on-arc-agi-3-demonstrating-a-frontier-level-general-purpose-architecture-for-long-horizon-autonomous-agents/

ARC Prize — Claude Opus 5 ARC-AGI results: https://arcprize.org/results/anthropic-claude-opus-5

ARC Prize — ARC-AGI-3 benchmark: https://arcprize.org/arc-agi/3

ARC Prize — ARC-AGI-3 technical report: https://arcprize.org/media/ARC_AGI_3_Technical_Report.pdf