Nvidia Research Lifts Claude Opus 5 to 100% on ARC-AGI-3

Update (9 Nov 2026): Claude Fable 5.2 and Mythos 5.2 are out - prices, access and benchmarks are in my Claude Fable 5.2 guide.

👁 121 reads

Nvidia published research on Friday showing that Claude Opus 5 jumped from a 30% score to a perfect 100% on the ARC-AGI-3 interactive reasoning benchmark simply by running the model inside a custom-built agent harness - without any changes to the model itself.

ARC-AGI-3 is a benchmark built around 2D games with no instructions, requiring a model to independently figure out the rules and win. A 100% score means the model matches human performance. The 30% baseline was already the highest score recorded across all models Nvidia tested.

The critical addition was a supervisor component layered on top of the main working agent. Adel El Hallack, vice president of product in Nvidia’s AI unit, described it as acting “almost like a CEO to nudge the agent when it goes off direction or starts exploring a path that it might lead to a dead end, or re-explore a path that it had previously trod.” El Hallack also pushed back on how agents are typically understood: “Generally speaking, the world interprets an agent almost as an API of the model,” he said, adding that an agent is actually “the model. It is the scaffolding around the model, which we call the harness, i.e. the set of tools that it utilizes. It is the runtime and the associated skills and libraries that we give it access to.”

The result fits into a broader pattern. OpenAI, whose models scored less than 10% on ARC-AGI-3, conducted its own harness research last month and found that adjusting two settings tripled its models’ scores - though none reached 100%. Earlier, Microsoft tested 19 large language models on long-horizon document-editing tasks and found every model, including frontier ones, filled documents with errors.

For developers and builders working with AI agents today, the practical takeaway is direct: most current agent setups - Claude Code, Codex, Hermes and similar single-layer tools - leave significant performance on the table. Adding a supervisor layer that monitors and redirects the main agent is not a research novelty; it is a concrete architectural choice that, according to Nvidia’s results, can be the difference between a barely functional agent and one that completes long-horizon tasks reliably.

// read moreResearcher Breaks Claude Code Opus 5 Auto Mode 80% of the TimeNewsClaude Opus 5 System Prompt Reveals Export-Control SuspensionNewsGPT-6 Astra vs Claude Opus 5: price and context comparedBenchmarks

Kinetyra