Descriptions:
Sam Witteveen examines OpenAI’s “Open Responses” initiative—a proposed open API standard aimed at giving open-source models a common interface for agentic capabilities including tool calling, streaming, multimodal inputs, server-side hosted tools, and agentic loop handling. The video provides context for why OpenAI moved in this direction: the Claude API has become the dominant interface in the coding agent ecosystem, with Chinese model providers like Moonshot AI and ZhipuAI explicitly building Claude Code-compatible endpoints and marketing Claude Code subscriptions for their own models.
Witteveen walks through the technical specification in detail, highlighting support for reasoning token summaries, tool choice configuration, internally hosted tools (analogous to Google’s server-side sandbox and search tools), and structured agentic response handling. Early adopters include Hugging Face, Vercel, OpenRouter, LM Studio, Ollama, and vLLM—with the expectation that frameworks like SGLang and Google’s Gemma model line will follow. The video also includes code examples demonstrating the standard in practice.
A recurring theme is the competitive irony: Anthropic set the de facto API standard through organic Claude Code adoption rather than a formal standardization push, forcing OpenAI to pursue community buy-in. Witteveen closes with a cautionary note about the direction model providers are heading in 2026—increasingly running agentic logic server-side, reducing developer visibility into what the model is actually doing—making the choice of API standard a consequential architectural decision.
📺 Source: Sam Witteveen · Published January 20, 2026
🏷️ Format: Deep Dive
