NVIDIA’s AVO agent system scored a perfect 100.00 on the ARC-AGI-3 public benchmark on August 21, 2026, solving all 183 levels across 25 game environments. The underlying model — Anthropic’s Claude Opus 5 — scores approximately 30 per cent on the same benchmark when run without the agent harness. The result is the clearest demonstration yet that agent architecture, not raw model capability, is becoming the decisive variable in long-horizon AI tasks.
What is ARC-AGI-3? It is an interactive reasoning benchmark where an AI agent enters unfamiliar game-like environments with no instructions, no stated rules, and no stated goals. The agent must explore through trial and error, infer the environment’s dynamics, discover objectives, and plan actions efficiently across progressively harder levels. Performance is measured using Relative Human Action Efficiency (RHAE), which combines task completion with per-level action efficiency compared to first-time human baselines. When we covered ARC-AGI-3 in April, the best models scored under 1 per cent. Humans average 80 per cent.
What AVO Actually Is
AVO — Agentic Variation Operators — is a general-purpose coding agent system developed by NVIDIA’s research team. It was first demonstrated on GPU-kernel optimization tasks, where it ran continuously for seven days, explored more than 500 optimization directions, and produced 40 committed kernel versions. On NVIDIA DGX B200 systems, the resulting attention kernels outperformed cuDNN by up to 3.5 per cent and FlashAttention-4 by up to 10.5 per cent.
The architecture has two mechanisms that distinguish it from standard agent loops. Persistent memory carries forward prior implementations, evaluation results, and accumulated reasoning, so the agent resumes from its current state rather than rebuilding context each time. A supervisor monitors the broader trajectory for stagnation — the usual failure mode where an agent loops on a strategy that stopped working — and redirects the main agent toward alternative approaches when progress plateed.
The stagnation detector is the design element worth paying attention to. During the seven-day kernel run, the main agent decided what to inspect, change, test, and evaluate, while the supervisor intervened only when the search stalled. That division of labour is what kept the loop productive over hundreds of iterations without human input.
The 30 Per Cent to 100 Per Cent Gap
The headline number is the gap. Claude Opus 5 alone scores roughly 30 per cent on ARC-AGI-3 at High reasoning effort, according to ARC Prize’s own results page. Wrapped in AVO, the same model cleared all 183 levels with a 100.00 RHAE score, using 6,624 environment actions. For comparison, VISTA — a competing agent system from independent research — used 7,542 actions with the same model to complete the same levels. AVO used approximately 12 per cent fewer actions.
NVIDIA was careful about what this comparison means. The two systems differ in agent backend, observation representation, memory, context management, and other implementation details. This is not a controlled ablation. AVO used a text-only 64x64 grid representation; VISTA’s primary configuration uses a rendered 512x512 PNG image. The memory system, designed to carry useful understanding forward and reduce repeated exploration, may matter more over long horizons than any single design choice.
The point NVIDIA is making — and it is a fair one — is that evaluating a model is not the same as evaluating an agent. Model capability matters enormously, but the surrounding system determines how effectively that capability converts into sustained autonomous progress.
Why This Matters Beyond Benchmarks
The practical implication is that the frontier of useful AI capability is moving faster than model releases suggest. If a 30-per-cent model becomes a 100-per-cent agent through better scaffolding, then the bottleneck is not waiting for GPT-6 or Claude Fable 6 — it is building better agent architectures. The same underlying pattern showed up in NVIDIA’s kernel optimization work: AVO operated autonomously for a week and beat hand-tuned library implementations. The domain changed from code optimisation to interactive reasoning, but the core agent loop — hypothesise, act, observe, revise, preserve useful state, recover from incorrect assumptions — did not.
For the broader industry, this reframes the agent architecture conversation. Companies investing in agent harnesses, memory systems, and supervision loops may be buying more real-world capability than those waiting for the next model release. The model is the engine. The harness is the transmission, suspension, and steering. A perfect engine with no transmission goes nowhere.
The Saturation Question
Saturating a benchmark is a mixed signal. A perfect 100 means ARC-AGI-3 can no longer distinguish between systems at the top — the benchmark is solved for the public set. The private holdout set, which ARC Prize uses for official competition scoring, remains untouched. Whether AVO generalises to the private set, or to genuinely novel environments outside the benchmark distribution, is the open question.
The earlier SeedIQ active inference result — 95 per cent on ARC-AGI-3 using a fundamentally different architecture — already signalled that the benchmark was approachable. NVIDIA’s perfect score confirms it. The next benchmark will need to be harder, or structured differently, to continue separating systems.
❓ FAQ
Does this mean Claude Opus 5 is now an AGI?
No. It means Claude Opus 5, when wrapped in a sophisticated agent system with persistent memory and stagnation detection, can solve a specific class of interactive reasoning puzzles. ARC-AGI-3 tests one dimension of fluid intelligence. Real-world generalisation remains an open question.
What is RHAE?
Relative Human Action Efficiency. It measures how efficiently an agent completes a task compared to a first-time human baseline, combining task completion with action economy. A score of 100 means the agent matched or exceeded human efficiency across all levels.
Is the 100 score on the public or private set?
Public set only — all 183 levels across 25 environments. The private holdout set used for the official ARC Prize competition was not tested.
Can I use AVO?
AVO is a research project. NVIDIA has not announced a commercial release. The architecture is described in detail in an arXiv paper, and the core design principles — persistent memory, stagnation detection, hypothesis-act-observe-revise loops — can be implemented with existing agent frameworks.
🔍 THE BOTTOM LINE
NVIDIA’s AVO result is a signal that the agent architecture layer is maturing faster than the model layer. The same model that scores 30 per cent alone scored 100 per cent wrapped in the right harness. That matters because it means the path to more capable AI systems runs through better scaffolding, not just bigger models. The benchmark is saturated for the public set, which means the field needs harder tests. But the lesson — that the harness does the heavy lifting — will outlast any single benchmark.
