Nvidia's Agentic Variation Operators (AVO) system boosted Claude Opus 5 from a baseline 30.2% to a perfect 100% on the ARC-AGI-3 advanced reasoning benchmark, according to a team blog released Friday by five Nvidia engineers and research specialists. The result demonstrates that architectural design surrounding an AI model — not just the model's raw power — can unlock what the team calls "frontier-level long-horizon performance." First unveiled in late March 2026, AVO functions as a general-purpose coding agent that handles tasks like inspecting and editing code, executing commands, consulting documentation, and validating outputs through execution, with a core emphasis on sustained autonomous operation across multistep challenges.
The AVO system achieved a 100.00 Relative Human Action Efficiency (RHAE) score across all 25 environments in the ARC-AGI-3 public set, completing all 183 levels. The ARC-AGI-3 benchmark employs RHAE, a measure that blends task completion with per-level action efficiency compared to first-time human baselines, with performance aggregated across different levels and environments. The 30.2% baseline for Claude Opus 5 at high reasoning effort on the public set was reported this July by ARC Prize, the nonprofit AI research and benchmarking organization. ARC Prize noted that Claude Opus 5 at maximum reasoning effort scored 97.5% on ARC-AGI-1 and 90.4% on ARC-AGI-2 semi-private, demonstrating strong logical reasoning. Nvidia additionally tested AVO paired with GPT-5.6 Sol on a challenging subset of games, where Sol reached matched levels faster in real-world time in several cases, while Opus used fewer environment actions in matched-level comparisons.
"Model capability matters enormously, but the surrounding system determines how effectively that capability can be converted into sustained autonomous progress," wrote the Nvidia blog team, led by principal engineer Terry Chen. The team emphasized that evaluating a model isn't the same as evaluating an agent, stating the result "illustrates a broader point" about the distinction between raw model capability and system-level performance. The assembled engineers said the most important result wasn't simply the perfect score, but the fact that the same agent architecture transferred from highly specialized GPU-kernel optimization to a very different interactive reasoning task. The team's final words restated the investigation's core finding: the model matters, but the model isn't the entire agent.
AVO replaces what Chen and colleagues describe as the "predefined variation step of conventional evolutionary-search systems" with an autonomous agent that decides how to generate the next candidate — determining what to inspect, what to change, what to test, and what to commit. The system requires two specific mechanisms to sustain long-horizon tasks beyond a single model's context window: persistent memory and supervision. Persistent memory carries forward prior implementations, evaluation results, compiler and profiler outputs, and accumulated reasoning, allowing the agent to resume from its current state instead of repeatedly reconstructing the search. Supervision involves a programmatic software module within AVO that monitors the broader search trajectory and can intervene when progress stalls. The team noted that GPU-kernel optimization and ARC-AGI-3 appear very different on the surface — one involves source code, compilers, profilers, and throughput, while the other involves unfamiliar interactive environments where an agent must infer the effects of available actions, discover objectives, and act efficiently. But the underlying computational pattern is similar, requiring the agent to build hypotheses from incomplete evidence, take actions through an external interface, observe consequences, preserve useful state, revise its problem model, recover from incorrect assumptions, and continue making progress over a long horizon.
The team concluded that generality can come not only from domain knowledge but from the machinery that allows reasoning and feedback to compound over time, with long-horizon capability emerging as a property of the full system. Memory determines what survives, tools determine what actions are possible, feedback grounds progress, and recovery allows work to continue beyond a single model invocation, the report states. The results cover the 25-environment ARC-AGI-3 public set using the official scorecard and RHAE metric, not results on the semi-private or fully private competition sets. The Nvidia team included Chen, director and distinguished engineer in competitive machine learning Jean-Francois Puget, VP of high-performance AI Humphrey Shi, and AI research interns Yeyin (Eva) Zhu and Zhifan Ye. Organizations investing in autonomous AI capabilities may need to balance model selection with equal attention to the orchestration layer that converts computational potential into sustained execution. The implication for enterprises is that proprietary advantage in AI deployment might increasingly reside in system architecture rather than model access alone.