Nvidia says the harness, not the model, took Claude Opus 5 from 30.2% to 100% on ARC-AGI-3

Nvidia has published the architecture of Agentic Variation Operators (AVO), the general-purpose coding agent system it introduced in late March 2026, and applied it to ARC-AGI-3 [s1 c1]. In a blog posted Friday by a five-person team of engineers, machine learning specialists and research interns led by principal engineer Terry Chen, the company reports that Claude Opus 5 went from a 30.2% baseline on the benchmark to a 100.00 RHAE score, with the harness as the only thing that changed ^[10][2][3].

The numbers first. ARC Prize, the non-profit that maintains the benchmark, reported 30.2% for Claude Opus 5 at high reasoning effort on the ARC-AGI-3 public set in an analysis post this July ^[2]. Nvidia reports 100.00 RHAE across all 25 environments in that public set, with all 183 levels completed ^[3]. That is a gap of 69.8 points, or roughly 3.3 times the baseline ^[22]. RHAE combines task completion with per-level action efficiency measured against first-time human baselines, aggregated across levels and environments ^[12], so it is not a simple pass rate and a perfect score means the agent was also action-efficient.

What makes this worth reading rather than filing is the claim about where the capability came from. Nvidia's framing is that the harness determines how a model receives context, uses tools, maintains state, responds to feedback, recovers from failure and sustains progress on long-running work ^[21]. The team writes that the result shows system design, not model capability alone, can unlock frontier-level long-horizon performance, and that "model capability matters enormously, but the surrounding system determines how effectively that capability can be converted into sustained autonomous progress" ^[20][19].

The transfer story is the part operators should look at. AVO was built for software engineering and GPU-kernel optimization, where it replaces the predefined variation step of conventional evolutionary search with an agent that decides what to inspect, change, test and commit ^[6]. For ARC-AGI-3, Nvidia says it connected the same agent to a different task interface, changing only the environment-specific tools and the evaluation ^[5]. ARC-AGI-3 drops agents into unfamiliar environments with no instructions, stated rules or stated goals ^[11], which on the surface has nothing to do with compilers and throughput, but Nvidia argues the loop is the same: build hypotheses from incomplete evidence, act through an interface, observe consequences, preserve state, revise, recover from wrong assumptions, keep going ^[13].

Two mechanisms carry that loop past a single context window: persistent memory and supervision ^[7]. Memory holds prior implementations, evaluation results, compiler and profiler output and accumulated reasoning so the agent resumes instead of rebuilding the search ^[9]. A supervisor watches the trajectory for stagnation and repeated unproductive cycles and can redirect the main agent ^[4]. In the kernel work, AVO ran continuously for seven days, explored more than 500 optimization directions and committed 40 kernel versions, with the supervisor stepping in when the search plateaued ^[15][8]. The resulting multihead attention kernels beat cuDNN by up to 3.5% and FlashAttention-4 by up to 10.5% on DGX B200 across the configurations evaluated ^[16], and the agent then adapted the kernel to grouped-query attention in about 30 minutes ^[17].

Two things to watch. The 100.00 is a public-set result ^[3]; ARC Prize's own comparative figures for Claude Opus 5 include a semi-private ARC-AGI-2 score of 90.4% ^[18], and neither Nvidia post reports a semi-private ARC-AGI-3 number. And the baseline and the agent result were produced by different parties, ARC Prize and Nvidia respectively ^[2][3], which is the usual reason to wait for a third-party run before treating a 69.8-point delta as a property of the harness rather than of the setup.