OpenAI adds WebSocket mode to Responses API

This is the kind of plumbing upgrade that matters more than a flashy model bump for agent builders. OpenAI is turning the Responses API into something that behaves more like a real session transport for tool-heavy workflows, not just another stateless wrapper around inference.

• –Persistent sockets plus incremental inputs attack a real bottleneck: repeated context transfer across every tool call in an agent loop
• –The feature is squarely aimed at coding agents and orchestration systems, where latency compounds fast once a task fans out into dozens of model-tool turns
• –Compatibility with `store=false` and Zero Data Retention is a big deal for teams running agents over private code or sensitive internal workflows
• –The tradeoffs are real: one in-flight response per connection, no multiplexing, a 60-minute connection cap, and cache misses can still break continuation with `previous_response_not_found`
• –Net effect: agent frameworks built on OpenAI can lean less on custom transport hacks and more on the platform's native continuation model