Improved extraction of reasoning and cached tokens from OpenAI and DeepSeek responses (including streams). Ensured accurate cost calculation using registry metadata.