Nvidia's AVO turns Claude Opus 5's 30% into 100% on ARC-AGI-3

Anthropic's Claude Opus 5 did not get smarter between 21 August and the day before. What changed was the system wrapped around it, and that is the line Nvidia is selling to the market: the model is a commodity, the harness is where the value sits.

Nvidia researchers said their AVO (Agentic Variation Operators) system completed all 183 levels across all 25 public environments of ARC-AGI-3, an interactive benchmark that tests whether AI can learn new skills without being told the rules, according to The New Stack. ARC Prize had previously put Claude Opus 5 alone, at high reasoning effort, at 30.2% on the same public set. AVO adds a main agent, persistent memory, tools and a supervisor that intervenes when progress stalls — the same model, a different scaffold.

The efficiency numbers matter as much as the headline score. AVO needed 6,624 environment actions to clear the benchmark, 12% fewer than the 7,542 used by the rival VISTA harness. Nvidia's five-person team — Terry Chen, Yeyin Zhu, Zhifan Ye, Jean-Francois Puget and Humphrey Shi — also noted the unchanged agent had separately spent seven days optimising GPU kernels, beating FlashAttention-4 by up to 10.5% on DGX B200 hardware. AVO itself is a research project Nvidia introduced in March 2026, not a shipping product.

The result comes with real caveats. The public ARC-AGI-3 set is tutorial-like and had already been solved by other harnesses before AVO arrived. There is no private-set score, which is the test ARC's creator, François Chollet, treats as the genuine measure of generalisation. Nvidia itself said configuration differences between runs mean the 30-to-100 jump is not a clean measure of AVO's own contribution.

For Anthropic, the episode is double-edged. Claude Opus 5 is doing the reasoning that makes AVO's 100% possible, which keeps the model relevant to buyers shopping for agent infrastructure. But the fact that a third party's harness, not Anthropic's own tooling, produced the headline number hands pricing power to whoever builds the best scaffold — and that need not be Anthropic.

Nvidia's real audience here is enterprise buyers and hyperscalers deciding where to spend on agent infrastructure, not benchmark purists. A 12% cut in environment actions is an efficiency claim with a dollar figure behind it once scaled to production workloads. The gap left open — no private-set result, no proof AVO generalises beyond a known-solvable public set — is exactly what rivals building competing harnesses will target next.