AI Achievement: New Model Outperforms Human-Level Reasoning on Real-World Tasks
Key points Advances in artificial intelligence reach a milestone in 2025, epitomized by new models that can not only match but in some cases exceed human-level reasoning on real-world tasks. This milestone is changing the expectations of the capabilities of AI in various sectors.
Key Developments
OpenAI’s o3 Model Defeats ARC-AGI Benchmark
Under standard compute, OpenAI’s o3 model set a new state of the art with 75.7% performance on the ARC-AGI benchmark, while with high compute, this model achieved 87.5%— higher than the mean human score on this difficult test of abstract reasoning and adaptation 12.
The ARC-AGI challenge is set to evaluate whether an AI agent can solve new problems and show fluid human-like intelligence that transcends the particularities of specific tasks, and it involves reasoning about objects, space, and logical inference12.
This result is important in indicating that AI models are now able to generalize and reason outside their training data with such noisy labels, which is a crucial step towards artificial general intelligence (AGI)2.
Human-Level and Then Some: What the Tests Show
Even assuming a consensually agreed upon definition can be found, such measures and criteria for ‘what it means for human level capabilities to get solved’ are not universally accepted, and no such criterion is widely adopted among artificial intelligence researchers and the general people working on such problems [33].Multiple independent assessments between 2015 and 2025 show that new reasoning models running on different system environments, OpenAI [31, 32] (e.g., ChatGPT o3, o4-mini) and Google [33] (Gemini series) attain or surpass Turing-level of intelligence (outstanding performance at the average human for a broad set of reasoning model tasks)34.
In such specialist disciplines as legal reasoning these models are already showing Ph. es Graduate Level with all reasonable speed, and so exceed the performance of most people in some intellectual productions34.
The reasoning, adaptation, and problem solving of models can be applied not only to narrow, pre-determined tasks, but to open-ended and real-world35 phenomena.
Meta’s V-JEPA 2: Physical Reasoning and World Modeling388Physics In (deep) reinforcement learning, a physics engine can be used to simulate a virtual environment of a world with various entities.
And Meta’s V-JEPA 2 model pushes AI’s capabilities to comprehend and forecast on the physical interaction itself. That allows robots and AI agents to simulate, plan, and execute in the actual world.56
Learned from a data set of more than a million hours of video and down-sampled robotic experience, our model (V-JEPA 2) has the ability to make inferences about physical properties of objects, predict object behaviour, and generalize to novel scenarios and can adapt well even to unseen scenarios — with strong generalization and prediction accuracy 56.
Recent benchmarks proposed by Meta, including the IntPhys2 and the CausalVQA, provide further support for these capacities in cause-and-effect reasoning and physical plausibility56.
