OpenAI's general-purpose reasoning model disproves 80-year-old Erdős planar unit distance problem; Cohere releases Command A+ as fully open Apache 2.0 model

OpenAI announced that an internal model disproved a long-standing belief around the planar unit distance problem, a famous Erdős problem from 1946, discovering a new family of constructions that improves on square-grid-style solutions. The model was described as a general-purpose reasoning system rather than a domain-specific math solver, and produced approximately 125 pages of reasoning output.

Cohere released Command A+ as Apache 2.0 open weights, positioning it as its most powerful model yet and explicitly optimized for low hardware requirements. The model is roughly 218B MoE / 25B active, multimodal, supports 48 languages, and can run on relatively modest setups including as little as 2× H100s at W4A4. Artificial Analysis placed Command A+ at 37 on its Intelligence Index, with especially strong non-hallucination behavior.

Multiple new benchmarks revealed gaps in current AI systems: InferenceBench found frontier agents struggle with system-level engineering and dependency management; MINTEval evaluated long-context memory systems with average instance length of 138.8k tokens with up to 1.8M, yet average accuracy across 7 systems was only 27.9%, with the best at 33.4%. ThoughtTrace introduced a dataset of 10,174 thought annotations from 1,058 users across 20 models, reporting +41.7% gains for user behavior prediction and +25.6% for alignment.