OpenAI: gpt-oss-20b

OpenAI's gpt-oss-20b, an open-weight 21B parameter MoE model, consistently ranks among the fastest models and offers highly competitive pricing. Its architecture, with 3.6B active parameters per forward pass, is optimized for lower-latency inference and deployability on consumer or single-GPU hardware. The model demonstrates strong performance in several key areas. It achieves excellent accuracy in Ethics (99.0%), General Knowledge (99.0%), and Coding (92.0%), placing it in the 62nd, 71st, and 82nd percentiles respectively. Its Reasoning capabilities are also notable at 89.8% accuracy (80th percentile). However, the model exhibits significant variability in its Keyword Topic Relevance Classification performance, ranging from 0.0% to 100.0% accuracy across different instances of the benchmark. This suggests potential inconsistencies or sensitivity to specific test conditions within this category. A notable weakness is its performance in Mathematics, where it scores 66.7% accuracy (32nd percentile) with a very high duration of over 9 million milliseconds, indicating potential inefficiencies or struggles with complex mathematical problems. Hallucinations are moderately controlled at 70.0% accuracy (17th percentile), suggesting room for improvement in acknowledging uncertainty. The model's support for reasoning level configuration, fine-tuning, and agentic capabilities like function calling and tool use further enhance its versatility.