Nvidia researchers have built an agent system that pushed Anthropic's Claude Opus 5 to a perfect 100% score on ARC-AGI-3, one of the most demanding interactive reasoning benchmarks in AI — using the very same model that scores 30% when it runs on its own. The result, published Friday on the Nvidia Technical Blog, is the strongest evidence yet that the software wrapped around a frontier model matters as much as the model itself. For anyone tracking the latest AI developments, it marks a shift in where competitive advantage in agentic AI actually lives.
The system is called AVO, short for Agentic Variation Operators, and it is Nvidia's take on a general-purpose architecture for long-horizon autonomous agents — AI that can stay on task for hours or days instead of answering a single prompt. On the ARC-AGI-3 public set, AVO recorded a 100.00 RHAE score across all 25 game environments, completing all 183 levels in 6,624 environment actions using Claude Opus 5 as its backend model.
What ARC-AGI-3 Actually Tests
ARC-AGI-3 is an interactive reasoning benchmark created by the ARC Prize foundation. An agent is dropped into unfamiliar, game-like environments with no instructions, no stated rules, and no stated goal. It has to explore through trial and error, infer how the environment works, figure out what "winning" means, and then act efficiently enough to progress through progressively harder levels.
The benchmark's scoring metric, Relative Human Action Efficiency (RHAE), combines task completion with how few actions the agent wastes relative to first-time human players. That makes it a brutal test of sustained autonomy: the agent must remember what it learned, recover from mistakes, and budget its actions carefully across an entire run.
Nvidia's headline number gains context from a comparison. VISTA, a previous direct-interaction harness that also used Claude Opus 5, completed the same 183 public-set levels in 7,542 environment actions. AVO needed roughly 12% fewer actions to finish the job, according to Nvidia's blog post.
From GPU Kernels to Unknown Games
AVO was not built for puzzles. Nvidia first demonstrated the architecture on GPU-kernel optimization, a domain where small code changes can break correctness, memory behavior, and throughput in unpredictable ways. In one attention-kernel study, AVO ran continuously for seven days, explored more than 500 optimization directions, and committed 40 kernel versions. On NVIDIA DGX B200 systems, the resulting multihead attention kernels outperformed cuDNN by up to 3.5% and FlashAttention-4 by up to 10.5%. The agent then adapted its evolved kernel to grouped-query attention in about 30 minutes of additional autonomous work.
The same core loop — inspect, plan, implement, test, evaluate — was then pointed at ARC-AGI-3 with only the task interface changed. Two mechanisms carried over. The first is persistent memory, which stores prior implementations, evaluation results, compiler and profiler outputs, and accumulated reasoning, so the agent can resume from its current state rather than rebuilding context from scratch. The second is a supervisor, a second agent that watches the overall trajectory and intervenes when progress stalls.
The Supervisor Is the Breakthrough
TechCrunch, which interviewed Nvidia's Adel El Hallak, vice president of product in the company's AI unit, highlighted the supervisor as the most important ingredient. "It almost acts like a CEO to nudge the agent when it goes off direction or starts exploring a path that might lead to a dead end, or re-explore a path that it had previously trod," El Hallak told the publication.
The performance gap is stark. Nvidia's testing found that Claude Opus 5 without a sophisticated harness managed a 30% score on ARC-AGI-3 — the best result among the bare models tested, but nowhere near a full run. OpenAI's models have scored under 10% on the benchmark, and the company published its own research last month showing that tweaking just two harness settings tripled its scores. No model-only configuration has come close to the 100% that AVO achieved.
Notably, the AVO configuration ran in a text-only modality: each observation was supplied as an exact 64x64 text grid, with no images or image tokens sent to the model. The reasoning power came from the loop around the model, not from multimodal perception.
Why the Industry Is Betting on Harnesses
The finding lands amid a wave of evidence that agent infrastructure, not raw model capability, is the current bottleneck for autonomous AI. Databricks published benchmarking research in July showing that the harness dramatically impacts both performance and cost. "You can pick the same model but different harnesses, and you get significantly more cost if you use the wrong harness," Databricks CEO Ali Ghodsi told TechCrunch. "That itself can 2x your cost."
El Hallak framed Nvidia's broader argument in terms of openness. "We believe in having an open agent stack — where you have control across the harness, across the infrastructure, across the runtime — is what's required for us to usher the ecosystem forward and securely," he said. Nvidia has open-sized pieces of harness technology under its NeMo brand, and the AVO research is documented in a paper on arXiv.
A Benchmark With a History
ARC-AGI-3 has become a flashpoint in the frontier-lab rivalry. OpenAI previously claimed strong results for its GPT-5.6 Sol model on the benchmark, drawing scrutiny over the evaluation settings used. Nvidia's result sidesteps that debate in one sense — it does not claim a smarter model, only a better-engineered agent around an existing one.
The message for developers and enterprises building agentic products is direct: model selection still matters, but system design now decides whether a frontier model delivers frontier results. As Nvidia puts it, evaluating a model is not the same as evaluating an agent — and the benchmark tables of 2026 are increasingly rewarding the engineers who build the scaffolding.
Stay Ahead of AI
Agent architectures are moving faster than ever, and the gap between a good harness and a bad one is now measured in benchmark points and real dollars. Follow AI Buzz Wire for daily, source-verified coverage of AI research, benchmarks, and product launches.
Read more AI research news →
