SAN FRANCISCO — OpenAI on Wednesday released an upgraded version of its flagship language model with native autonomous agent features designed to execute multi-step workplace tasks without continuous human prompting. The company published benchmarks alongside the launch.

The release deepened a market-wide pivot toward AI agents—software that can plan and carry out sequences of actions across applications. Companies including Fireflies.ai have already recast assistants as workplace agents, with Fireflies this week expanding its Fred agent in markets such as India. OpenAI's move brings agentic capability directly into its core model rather than as a separate layer.

According to documentation on OpenAI's website, the model improved reliability on long-horizon tasks and reduced errors during tool use. The launch included expanded enterprise controls, permissioning for agent actions, and audit logging aimed at corporate customers wary of granting autonomous software access to internal systems.

The timing placed additional pressure on rivals Anthropic and Google, both of which have prioritised agentic products for business clients. Security researchers have warned that more capable agents raise fresh risks, echoing concerns raised this week over AI-assisted hacking targeting hospitals and banks. OpenAI said new safeguards restrict high-risk actions by default.

The release arrived days after US senators pressed OpenAI and Anthropic on copyright and child safety, underscoring the regulatory scrutiny now shadowing each major model launch.