OpenAI has introduced ChatGPT Work, a significant departure from its chat-first positioning. Rather than requiring users to toggle between applications and prompts, ChatGPT Work operates as an autonomous agent capable of taking actions across integrated apps and files while maintaining context over multi-hour project sessions. The agent can navigate Word documents, pull data from Excel, modify presentations, and coordinate across tools—automating what would traditionally require manual context-switching. This capability represents OpenAI's implicit acknowledgment that enterprise customers increasingly demand agents that reduce friction across their existing software stacks, not just better conversation partners. The timing matters: as Microsoft deepens ChatGPT integration into Microsoft 365 Copilot using the new GPT-5.6 model, OpenAI is simultaneously launching a competing agent product that threatens to create switching costs independent of Microsoft's ecosystem. This suggests OpenAI is hedging against deeper Microsoft control of the enterprise distribution channel.
GPT-5.6 is the engine behind both Microsoft 365 Copilot's upgrade and ChatGPT Work. According to OpenAI's positioning, GPT-5.6 delivers stronger performance per token and lower operational costs compared to its predecessor, a critical metric in enterprise sales where per-user economics determine adoption curves. The real competitive signal: if GPT-5.6 genuinely improves cost-per-token while maintaining quality, OpenAI undercuts rivals like Anthropic's Claude and Google's Gemini on unit economics—the hidden battleground in enterprise AI where customers are hyper-sensitive to recurring costs. Deutsche Telekom's deployment of OpenAI capabilities across customer service, employee workflows, and network operations demonstrates early validation that enterprises will embed these models into production systems, not just experimentation sandboxes. However, specific deployment metrics remain opaque. How many customer service interactions? What's the productivity gain? Are there failure modes in network operations where AI errors carry operational risk?
The strategic implications are substantial. By launching ChatGPT Work as a persistent, multi-app agent, OpenAI is shifting competition away from model capability (where margins compress) toward platform stickiness and workflow integration depth. If ChatGPT Work can sustain a complex project—say, a user briefing it to build a quarterly report requiring data pulls, analysis, writing, and design across multiple tools—for hours without user intervention, it becomes operationally embedded in enterprise processes. This creates customer lock-in independent of model superiority alone. Claude and Gemini must now match not just chat performance but agent reliability and app integration breadth. OpenAI's move also signals confidence in GPT-5.6's cost structure and reasoning capabilities; agents fail catastrophically if the underlying model hallucinates or makes costly errors during unsupervised execution. The Deutsche Telekom partnership validates demand, but enterprises will demand concrete SLAs and failure analysis before deploying agents into mission-critical workflows. OpenAI's ability to deliver production-grade reliability at scale will determine whether ChatGPT Work becomes a platform lock-in tool or remains a productivity demo.