ChatGPT Work launches GPT‑5.6 agents for multi‑step office tasks

ChatGPT Work is a clear marker of where frontier labs think the next leg of AI competition will play out: not just in raw model quality, but in end‑to‑end agents that can quietly absorb whole workflows. By bundling GPT‑5.6 with Codex inside a persistent “Work” mode, OpenAI is explicitly targeting the everyday projects that occupy analysts, product managers and operations teams—turning ChatGPT from a Q&A interface into a digital colleague that can stay on a task for hours.([s2pmag.ch](https://s2pmag.ch/chatgpt-work-openai-lance-un-agent-ia-capable-de-travailler-a-votre-place-grace-a-gpt-5-6/))

Strategically, that shifts the battleground with Anthropic’s Claude Cowork and similar offerings from Google and Microsoft into the heart of enterprise productivity suites. Whoever owns the agent layer that orchestrates files, tools and approvals will control far more value than a generic API provider. It also creates a rich feedback loop for OpenAI: the more real‑world multi‑step tasks these agents run, the more data they have about failure modes, reasoning chains and human override patterns—exactly the ingredients needed to train more capable, more aligned successors.([s2pmag.ch](https://s2pmag.ch/chatgpt-work-openai-lance-un-agent-ia-capable-de-travailler-a-votre-place-grace-a-gpt-5-6/))

For the race to AGI, this is a step toward systems that don’t just respond, but pursue goals over time with increasing autonomy. The big open questions now are governance and reliability at scale: how often these agents silently break, how organizations rein them in, and whether workers are empowered or just monitored more efficiently.