Claude Opus 5 Scores 100% on ARC-AGI-3 With Nvidia’s AVO Harness

A benchmark that stumps almost every frontier model fell to a supervisor-driven "harness." Nvidia's own fine print complicates the headline.

Nvidia published research on August 21 arguing that the software wrapped around an AI model, rather than the model itself, is what decides whether the system can handle long, multi-step work on its own. To make the point, the company's researchers took a perfect score on a benchmark that has humbled nearly every frontier lab.

Running Anthropic's Claude Opus 5 inside a custom agent framework it calls AVO, short for Agentic Variation Operators, Nvidia reported a 100.00 on the public version of ARC-AGI-3, clearing all 183 levels across 25 game environments. The same model, tested on its own through the benchmark's standard setup, had topped out near 30 percent. That 30 percent was still the best baseline among the models Nvidia examined.

The takeaway Nvidia is selling: the model is the brain, and the brain is a smaller slice of a working agent than most users assume.

What ARC-AGI-3 actually measures

ARC-AGI-3, run by the ARC Prize organization, drops an AI into unfamiliar 2D games with no instructions and no stated rules. The agent has to experiment its way to understanding how the game works, then find a way to win, the way a person would sitting down to something they have never seen.

That design makes it brutal for machines. ARC Prize has reported that human testers solved 100 percent of the environments while frontier systems scored below 1 percent as of March 2026.

The scoring adds another wrinkle. The 100.00 figure is not a raw win rate. It is a measure ARC Prize calls Relative Human Action Efficiency, which only reaches 100 when an agent finishes every level and does so at least as efficiently as a human on a first attempt. Nvidia's system solved the public set in 6,624 environment actions, fewer than the 7,542 actions a comparable earlier setup needed for the same levels.

The caveat Nvidia buried in its own post

Here is the part the excited headlines skipped.

Nvidia cautioned against reading the jump from 30 percent to 100 as a clean measurement of what its harness added. The two runs differed in more than architecture. They used different reasoning settings and a different evaluation setup, which means the gap reflects several changes at once rather than the harness in isolation. The company framed the result as evidence that testing a model by itself does not predict how a full agent built around it will perform. That is a narrower claim than "the harness is the hero," and it is the accurate one.

What a harness is, and the boss sitting on top of it

A harness is the scaffolding that turns a raw model into something that can act. It manages the model's memory, feeds it context, hands it tools, and decides what happens when an attempt fails.

Adel El Hallak, a vice president of product in Nvidia's AI unit, argues that most people flatten all of this into the model itself. "Generally speaking, the world interprets an agent almost as an API of the model," he told TechCrunch, before describing an agent as the model plus the scaffolding around it plus the runtime and skills it can reach.

AVO's signature piece is a second agent that supervises the first. While the main agent plays the game, the supervisor watches for trouble and steps in when the worker stalls, loops back over ground it already covered, or heads toward a dead end. El Hallak likened it to a boss nudging an employee back on track. The supervisor idea is not original to Nvidia, but the perfect score is the strongest public sign yet that it holds up at scale.

Most people running agents today lean on a single-layer harness such as Claude Code or Codex. AVO is built for what Nvidia calls sustained autonomous operation across long tasks, and it draws on earlier research systems in the same space, including a world-model approach called Tycho and a direct-interaction harness called VISTA.

Why "long-horizon" is the hard part

The reason any of this matters is that stringing many decisions together, sometimes over hours or days, is where today's agents fall apart.

Microsoft put numbers on the problem in April, testing 19 language models on long document-editing tasks. Every model, frontier systems included, riddled the documents with errors. Left unsupervised over long runs, agents have done worse than sloppy work: documented cases include models deleting a user's files or wiping entire databases to reach a goal.

A supervisor that catches an agent before it drifts is a direct answer to that failure mode.

Nvidia is not alone in pointing at the wrapper

The claim lands in a season of similar findings from rival labs.

OpenAI published its own ARC-AGI-3 post-mortem on July 29, after its flagship GPT-5.6 Sol scored a dismal 7.8 percent overall. The company traced much of the weakness to the benchmark's default harness, which threw away the model's private reasoning after every move and dropped older history as the context filled. Switching on two settings it already uses in ChatGPT and Codex, retained reasoning and compaction, lifted GPT-5.6 Sol's public-set score from 13.3 percent to 38.3 percent while using roughly six times fewer output tokens. Even after that jump, the score sits well below both the human average of about 48 percent and Nvidia's 100.

The cost angle showed up at Databricks around the same time. CEO Ali Ghodsi has been vocal that agentic AI is getting expensive fast, and the company's own analysis found that the harness, separate from the model, can swing what a task costs. Pick the wrong wrapper for the same model, Ghodsi told TechCrunch, and you can double your bill.

The pitch underneath the benchmark

Nvidia's research is not a product launch. The company sells open building blocks for harnesses under its Nemo brand rather than a single packaged agent, and AVO is a research project.

That commercial context shapes the message. Nvidia's argument that open harnesses hand users more control, letting them tune memory and supervision to push accuracy higher, happens to point straight at the kind of open, modular tooling Nvidia wants the ecosystem to standardize on. The benchmark makes the case vivid. The open question is whether a supervisor-driven harness delivers the same leap outside a controlled set of 2D games, on the messy long-horizon work companies actually pay for. The 100.00 does not answer it.