OpenAI's GPT-OSS models benchmarks worse than DeepSeek R1 and Qwen3 235Bxcancel.com8 pointspu_pea year ago