# NVIDIA's AVO scores 100% on ARC-AGI-3's public set > Source: > Published: 2026-08-21 18:26:59+00:00 # NVIDIA's AVO scores 100% on ARC-AGI-3's public set **Powered by Claude Opus 5, the coding-agent system cleared 183 levels in 6,624 actions, though NVIDIA did not test the private sets.** By [Ryan Merket](/author/ryan-merket) ยท Published Primary source: [NVIDIA AI on X](https://x.com/NVIDIAAI/status/2090786258981466231) ## Why it matters AVO shows how much agent memory, tools and supervision can change the performance of the same underlying model. Its 100% score is limited to a public set already solved by other harnesses, leaving private-set generalization as the harder test. [NVIDIA AI (@NVIDIAAI)](https://x.com/NVIDIAAI/status/2090786258981466231) said Friday that its AVO coding-agent system scored 100% on the public ARC-AGI-3 benchmark, completing all 183 levels across 25 interactive environments without instructions, stated rules or defined goals. [https://x.com/NVIDIAAI/status/2090786258981466231](https://x.com/NVIDIAAI/status/2090786258981466231) The result came from a research group including [Terry Chen](https://developer.nvidia.com/blog/author/tairuic/), Yeyin Zhu, Zhifan Ye, Jean-Francois Puget and [Humphrey Shi (@humphrey_shi)](https://x.com/humphrey_shi/with_replies). Chen, an NVIDIA principal engineer and former vice president of engineering at HippoML, helped build the AITemplate GPU optimization framework. Shi leads NVIDIA's High-Performance AI group, which works across models, kernels, compilers and hardware. In a [technical post published alongside the result](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/), the researchers identified the underlying model as [Claude Opus 5](/models/azure/claude-opus-5). AVO supplied the surrounding agent system: persistent memory, software tools, an execution loop and a supervisor that can redirect work when progress stalls. ### What the perfect score measures [ARC-AGI-3](https://arcprize.org/blog/arc-agi-3-launch) presents agents with unfamiliar, turn-based environments. An agent has to experiment, infer what its available actions do, discover the objective and carry that knowledge into later levels. The benchmark launched in March with humans scoring 100% and frontier AI systems at 0.51%, according to the ARC Prize Foundation. Its [Relative Human Action Efficiency metric](https://docs.arcprize.org/methodology) rewards both completion and efficient play. A 100% score means a system completed every level while matching or exceeding the human action baseline. That metric counts interactions that change the game state. Internal reasoning, tool calls and retries do not count as actions. Agent developers can therefore spend substantial inference and computation between moves without lowering the score, provided the resulting system acts efficiently inside the environment. NVIDIA said AVO completed the public set in 6,624 environment actions. The researchers compared that with 7,542 actions reported by [VISTA](https://vista-research.github.io/), a separate Claude Opus 5 harness from MIT researchers that also reached 100%. NVIDIA calculated that AVO used about 12% fewer actions, while warning that the comparison was not a controlled test because the systems use different memory designs, observation formats and agent backends. AVO read each observation as an exact 64 x 64 text grid. It received no images or image tokens. VISTA's primary configuration instead feeds Claude a rendered 512 x 512 PNG and retains earlier visual observations for later retrieval. ### A public benchmark that agents have already saturated AVO is not the first system to post a perfect score on the public environments. The [ARC Prize community leaderboard](https://arcprize.org/leaderboard/community) listed Tycho at 100% on July 29th, and VISTA published its perfect result on August 5th. NVIDIA's run also did not cover ARC-AGI-3's semi-private or fully private competition sets. Those withheld environments are the stronger test of whether an agent has learned a transferable method rather than accumulated useful knowledge about publicly available games. The public score still provides evidence for NVIDIA's narrower claim: a coding-agent architecture can transfer beyond conventional software tasks. AVO used the same basic loop for ARC-AGI-3 that NVIDIA previously applied to GPU-kernel optimization, with environment-specific tools replacing compilers, profilers and code tests. ### From GPU kernels to interactive games NVIDIA introduced AVO in a [March research paper](https://arxiv.org/abs/2603.24517) as an autonomous variation operator for evolutionary search. Instead of asking a language model to generate one candidate inside a fixed optimization pipeline, AVO lets the agent inspect prior attempts, edit code, execute tests, interpret results and decide what to try next. In NVIDIA's attention-kernel experiment, AVO ran for seven days, explored more than 500 optimization directions and committed 40 kernel versions. The researchers reported gains of up to 3.5% over cuDNN and 10.5% over FlashAttention-4 on the tested NVIDIA B200 configurations. AVO then adapted the resulting work to grouped-query attention in roughly 30 minutes. The ARC-AGI-3 result advances the same system-level argument. Claude Opus 5 supplied the base model capability, while AVO preserved state, managed feedback and sustained work beyond a single context window. ARC Prize separately reports a score of roughly 30% for Claude Opus 5 under its own public-set evaluation, although NVIDIA cautioned that different reasoning settings and evaluation setups prevent a clean attribution of the improvement to AVO alone. For NVIDIA, AVO extends the company's position beyond providing the chips that run frontier models. The research places NVIDIA in the agent-control layer, where memory, tools, supervision and execution determine how much useful work a model can complete. The perfect public score shows that layer can dominate benchmark performance. The private environments will determine how far that performance travels.