Nvidia announced that Claude Opus 5 achieved a 100% score on the ARC-AGI-3 test thanks to its software stack
Nvidia released a study in which it stated that specially tailored software for an AI agent enabled the Claude Opus 5 model to score 100% on the ARC-AGI-3 interactive reasoning test. Without this framework, the model scored 30%—the highest score among the models tested, according to TechCrunch.
Nvidia defines the framework as a set of components surrounding the language model that manage memory, context, and feedback. It also includes tools, a runtime environment, as well as skills and libraries available to the agent. According to the company’s researchers, the choice of model is important, but in tasks requiring a large number of sequential decisions, the result also depends significantly on this infrastructure.
ARC-AGI-3 consists of two-dimensional games without instructions. The model must independently determine the rules, understand how to play, and win the games. According to TechCrunch, a score of 100% means that the model performs at a human level in these games.
For more breaking news, follow the UA.News Telegram channel.
For the tests, Nvidia used its own framework, Agentic Variation Operators (AVO). It includes a separate supervisory agent in addition to the main one: this agent guides the main agent if it reaches a dead end or begins to re-explore a path it has already taken. Adel El Hallak, vice president of products for Nvidia’s AI division, told TechCrunch that the agent is not limited to the model’s API but encompasses the model itself and the system of components surrounding it.
Nvidia did not present AVO as a standalone product. The company is developing technologies to create such frameworks under the NeMo brand; some of these tools are openly available, while others are commercial.
Previously, OpenAI also investigated the performance of its models in ARC-AGI-3. According to TechCrunch, changing two parameters of the framework allowed them to triple their performance, though none of the tested models reached 100%. A separate study by Databricks, published in July, showed that different architectures for the same model can significantly impact the cost of running AI agents.
