The new OpenAI Agents Platform
OpenAI introduced a comprehensive suite of new tools for AI agents, including the Responses API, Web Search Tool, Computer Use Tool, File Search Tool, and an open-source Agents SDK with integrated observability tools, marking a significant step towards the "Year of Agents." Meanwhile, Reka AI open-sourced Reka Flash 3, a 21B parameter reasoning model that outperforms o1-mini and powers their Nexus platform, with weights available on Hugging Face. The OlympicCoder series surpassed Claude 3.7 Sonnet and much larger models on competitive coding benchmarks. DeepSeek built a 32K GPU cluster capable of training V3-level models in under a week and is exploring AI distillation. Hugging Face announced Cerebras inference support, achieving over 2,000 tokens/s on Llama 3.3 70B, 70x faster than leading GPUs. Reka's Sonic-2 voice AI model delivers 40ms latency via the Together API. Alibaba's Qwen Chat enhanced its multimodal interface with video understanding up to 500MB, voice-to-text, guest mode, and expanded file uploads. *Sama* praised OpenAI's new API as "one of the most well-designed and useful APIs ever."