AI Daily — April 23, 2026
2026-04-23
openai.com
Introducing workspace agents in ChatGPT
OpenAI has launched Workspace Agents in ChatGPT — Codex-powered cloud agents that automate complex, multi-step workflows across connected tools for enterprise teams. Agents run persistently in the cloud, can be scoped to team contexts, and integrate with external services securely. This represents a significant step toward production-grade agentic automation within ChatGPT's enterprise offering.
openai.com
Speeding up agentic workflows with WebSockets in the Responses API
OpenAI published a technical deep-dive on the Codex agent loop, detailing how WebSocket connections and connection-scoped KV caching were used to significantly reduce per-turn API overhead in multi-step agentic workloads. The approach avoids repeated HTTP handshakes and enables prompt cache reuse across turns in the same session. This is directly actionable for developers building latency-sensitive agentic applications on the Responses API.
openai.com
Introducing OpenAI Privacy Filter
OpenAI released an open-weight model specifically trained for PII detection and redaction in text, claiming state-of-the-art accuracy across standard benchmarks. The model is designed to be integrated into data pipelines as a preprocessing step, particularly relevant for enterprise and healthcare deployments. Releasing weights is notable for OpenAI and makes the model auditable and self-hostable.
openai.com
Making ChatGPT better for clinicians
OpenAI is offering ChatGPT for Clinicians free to verified U.S. physicians, nurse practitioners, and pharmacists, targeting clinical documentation, decision support, and research use cases. The move follows a competitive push in healthcare AI and lowers the barrier for professional adoption significantly. Verification of clinical credentials is required to access the offering.
rss.arxiv.org
Expert Upcycling: Shifting the Compute-Efficient Frontier of Mixture-of-Experts
This paper proposes a method to progressively expand MoE model capacity by increasing expert count during continued pre-training rather than training from scratch, constructing an mE-expert model from a trained E-expert checkpoint. The approach addresses the high memory and communication costs of training large MoEs by reusing existing weights. Results suggest this shifts the compute-efficient Pareto frontier, making larger MoEs accessible at lower training budgets.