ChatGPT Work vs Claude Cowork: The AI War Moves From Smartest Model to Cheapest Agent

OpenAI unveiled ChatGPT Work, an agent that fuses its chatbot with its Codex coding tool so that white-collar workers who can't code can still generate documents, presentations and websites. It runs on GPT-5.6, launched the same day in three sizes — a model whose release, the company notes, was delayed last month at the U.S. government's request over national-security concerns. The framing is explicitly competitive: ChatGPT Work is a direct answer to Anthropic's Claude Cowork, the autonomous multi-step agent Anthropic shipped in January, and it also squares up against Microsoft's Copilot Cowork. OpenAI stressed its offering is cheaper and more broadly available, rolling out first to Pro, Enterprise and Edu users before reaching Plus and Business.

The strategic logic is clear once you follow the money. Both OpenAI and Anthropic are reportedly preparing for possible public offerings, and both are chasing enterprise business, which is far more lucrative than selling to consumers. The prize is the professional who wants coding-tool power without the specialized knowledge — a huge, underserved middle. Whoever owns that workflow owns the recurring enterprise revenue, and that is what this launch is really about.

But the detail worth pausing on isn't the flagship. According to Creative Strategies analyst Max Weinbach, the smallest version of the new model can complete a task about as well as the largest — at roughly one-fifth the cost — and he calls it the first time he's seen small models handle these kinds of tasks. OpenAI's own product manager describes the 5.6 line as competitive with models that are "far, far more expensive," at twice the speed and much cheaper. That is the story hiding under the super-app headline: the industry is anxious about the high cost of running these tools, and the frontier is quietly shifting from "most capable" to "good enough, at a price that scales."

Our reading: this is the maturation we've been tracking, where the battle moves away from the smartest single model and toward distribution, cost, and the plumbing that puts agents into real workflows. When a fifth-the-cost model does the job, capability stops being the moat and economics take over — which is broadly good news, because collapsing inference costs are what turn a lab demo into abundance available to ordinary teams. Two cautions temper the optimism. First, treat "competitive with far more expensive models" as vendor framing until independent benchmarks confirm it — launch-day performance claims are marketing until measured. Second, the government-requested delay over national-security fears is a quiet but important marker: the most capable agents are now treated as strategic assets, not just products. The near term is a price war that democratizes access; the durable question is who controls the standards and rails these agents run on — and that contest is only beginning.