Key Takeaways
- NVIDIA’s AVO hit a perfect 100% RHAE on ARC-AGI-3 across all 25 public environments, using 12% fewer actions than VISTA.
- Persistent memory and a supervisory loop let AVO run for seven days on GPU kernels, then transfer to ARC-AGI-3 with zero instructions.
- AVO’s result — with Claude Opus 5 scoring ~30% standalone — shows agent architecture, not model alone, drives long-horizon success.
Table of Contents
A 100% Score That Reconfigures the Agent Stack
NVIDIA has pushed a single agent architecture to a perfect score on ARC-AGI-3, completing all 183 public-set levels across 25 unfamiliar interactive environments.
Reported by NVIDIA’s developer blog on August 21, 2026, the result records a 100.00 RHAE score while using roughly 12 percent fewer environment actions than the VISTA harness.
The more consequential signal is that the same general-purpose system had previously optimized GPU kernels for seven days without manual intervention, then transferred to a benchmark with no instructions, no rules, and no stated goals.
Inside the System-Level Machinery Behind AVO
According to NVIDIA’s developer blog, AVO, short for Agentic Variation Operators, is not a foundation model. It is a general-purpose coding agent system that decides what to inspect, change, test, and commit during long autonomous runs.
Two mechanisms distinguish it from single-context harnesses: persistent memory and a supervisory loop that detects stagnation and redirects the main agent.
Persistent memory preserves prior implementations, compiler and profiler outputs, and accumulated reasoning so the system can resume from its current state instead of reconstructing the search.
The supervisor monitors the broader trajectory and can intervene when unproductive cycles repeat, keeping forward progress from stalling.
NVIDIA first stressed this design on attention-kernel optimization, where AVO ran for seven days, explored more than 500 optimization directions, and committed 40 kernel versions.
On DGX B200 systems, the resulting multihead attention kernels outperformed cuDNN by up to 3.5 percent and FlashAttention-4 by up to 10.5 percent across evaluated configurations.
The agent later adapted the evolved kernel to grouped-query attention in roughly 30 minutes of additional autonomous work.
ARC-AGI-3 places agents inside unfamiliar game-like environments without instructions, explicit rules, or a stated goal.
Performance is measured by Relative Human Action Efficiency, or RHAE, which combines task completion with per-level action efficiency relative to first-time human baselines.
For the AVO configuration, the underlying model operated in text-only mode, receiving each observation as an exact 64 x 64 text grid with no image tokens.
NVIDIA’s engineering team reimplemented the task interface independently, drawing on direct-interaction principles described by VISTA rather than the programmatic world-model approach used by Tycho.
Across both GPU optimization and ARC-AGI-3, the core loop remains the same:
- Hypothesis formation from incomplete evidence
- Action execution through an external interface
- State preservation for accumulated learning
- Recovery from incorrect assumptions
The domain and feedback channel change, but the machinery for sustained autonomous progress does not.
What a Frontier Score Means for AGI Benchmarks and Market Stakes
Anthropic’s July 2026 announcement for Claude Opus 5 describes the model as scoring three times higher than the next-best model on ARC-AGI 3, but the company does not publish an absolute percentage.
That omission matters because NVIDIA’s 100.00 RHAE result was achieved with Claude Opus 5 running inside the AVO harness, not as a standalone model.
ARC Prize separately records approximately 30 percent for Claude Opus 5 at High reasoning effort on the public set, while AVO’s full system reaches 100.00 RHAE under a different reasoning setting and agent setup.
These figures should not be read as a controlled ablation, but they expose how much of a benchmark score belongs to the complete agent system rather than the model alone.
NVIDIA also paired AVO with GPT-5.6 Sol on a challenging subset of games, finding that Sol reached matched levels faster in wall-clock time in several cases while Opus used fewer environment actions.
These preliminary results suggest complementary operating profiles across frontier models.
IBM Think reports that researcher Ben Goertzel told AGI-26 in San Francisco that rapid progress in AI agents has increased the probability of an AGI breakthrough within the next 9 to 12 months.
The same report notes that no standard definition or accepted test of AGI exists, and some computer scientists reject the premise outright.
ARC-AGI-3 sits inside that unresolved debate: humans can solve all environments, yet an agent’s score depends on interface, memory, action efficiency, and recovery.
One constraint remains. NVIDIA’s perfect score covers the 25-environment public set, not the semi-private or fully private competition sets.
The model matters, but the model is not the entire agent.
Long-Horizon Agents Now Belong to the Architecture, Not Just the Model
The ARC-AGI-3 result shows that long-horizon capability is a property of memory, tools, feedback loops, and recovery — not raw model quality alone. For teams building autonomous agent pipelines that need to preserve state and scale efficiently, programmatic SEO and AI automation is how Andres SEO Expert approaches agent-driven workflows — contact Andres SEO Expert.
Frequently Asked Questions
What is AVO in NVIDIA’s ARC-AGI-3 announcement?
AVO, short for Agentic Variation Operators, is a general-purpose coding agent system developed by NVIDIA. It is not a foundation model but a system that decides what to inspect, change, test, and commit during long autonomous runs. It uses persistent memory and a supervisory loop to sustain progress.
How did NVIDIA’s AVO achieve a perfect score on ARC-AGI-3?
AVO scored 100.00 RHAE on the public set of ARC-AGI-3 by completing all 183 levels across 25 unfamiliar interactive environments. It used a text-only mode receiving 64×64 text grids and relied on the same core loop of hypothesis formation, action execution, state preservation, and recovery used in its GPU kernel optimization tasks.
What is RHAE in ARC-AGI-3?
RHAE, or Relative Human Action Efficiency, is the performance metric for ARC-AGI-3. It combines task completion with per-level action efficiency relative to first-time human baselines. AVO achieved a 100.00 RHAE score while using roughly 12 percent fewer environment actions than the VISTA harness.
What role does persistent memory play in AVO’s architecture?
Persistent memory preserves prior implementations, compiler and profiler outputs, and accumulated reasoning. This allows AVO to resume from its current state instead of reconstructing the search, enabling long-horizon autonomous operation across different domains without manual intervention.
How does AVO’s score compare to Claude Opus 5’s standalone performance on ARC-AGI-3?
According to the article, Claude Opus 5 running inside the AVO harness achieved the 100.00 RHAE result, while ARC Prize records approximately 30 percent for Claude Opus 5 at High reasoning effort on the public set. This shows that the full agent system contributes significantly to the score, not just the model.
Does NVIDIA’s 100% score on ARC-AGI-3 mean AGI has been achieved?
No. The perfect score covers the 25-environment public set, not the semi-private or fully private competition sets. Moreover, experts note that no standard definition or accepted test of AGI exists, and an agent’s score depends on interface, memory, action efficiency, and recovery.
What is the main takeaway from AVO’s performance for long-horizon agent design?
The result shows that long-horizon capability is a property of memory, tools, feedback loops, and recovery — not raw model quality alone. Agent architectures with persistent state and supervisory loops can transfer skills across domains like GPU kernel optimization and unfamiliar interactive environments.
