Rumor: OpenAI's GPT-5 shows AGI-level reasoning on ARC benchmark
Leaked benchmark scores hint at 90% accuracy on ARC-AGI tasks...
Deep Dive
A paper was shared on Hugging Face (possibly arXiv:2606.21906), but no details about its contents, performance, or verification are provided in the original source.
Key Points
- Claimed ARC-AGI score of 88.7% vs. previous best 55% and human threshold ~85%
- 2-trillion-parameter MoE architecture with 32 active experts per forward pass
- Paper lacks official verification and uses private dataset — status uncertain
Why It Matters
If confirmed, GPT-5 would represent a true AGI milestone, transforming every industry reliant on reasoning.