ChatGPT Work vs Claude Cowork: The Agentic Seat War
OpenAI and Anthropic are no longer fighting only on LMArena — ChatGPT Work and Claude Cowork compete for the surface where finished work ships.
The July 2026 fight between OpenAI and Anthropic is no longer about which model tops LMArena. It is about who owns the surface where finished work ships. On July 7, Anthropic took Claude Cowork to the web at claude.ai and to iOS and Android, with remote sessions that keep running when every device is offline. Two days later, OpenAI answered with ChatGPT Work — ChatGPT, Codex, and browser automation collapsed into one agent surface powered by the GPT-5.6 family.
The data Anthropic published with the Cowork expansion should end the "coding agents only" framing. From 1.2 million anonymized sessions across more than 600,000 organizations in a late-May sample, business process work took 33.4% of usage, content creation 16.4%, and software development only 8.7%. Over 90% of Cowork sessions are non-coding knowledge work: reports, onboarding checklists, decks, and spreadsheet reconciliation. OpenAI's ChatGPT Work aims at the same prize — finished documents, presentations, and web apps — and is rolling out first to Pro, Enterprise, and Edu users before Plus and Business.
Under the hood, helloai still tracks the models that power those surfaces. Claude Fable 5 leads text overall at 1508 Elo, Claude Opus 4.8 sits at 1503, and GPT-5.6 Sol is at 1484 with $5/$30 per million tokens. Those gaps matter for agentic coding and long-horizon runs. They matter less once a team standardizes on a single work surface that holds their files, context, and scheduled jobs. Product lock-in is becoming the durable moat as the model layer keeps leapfrogging week to week.
The trade-off is real. Cowork's remote sessions and Max-first mobile beta favor teams already deep in Anthropic's stack; ChatGPT Work's Codex-plus-browser pitch favors teams that want one super-app for artifacts. Neither product makes models interchangeable for free — routing a hard reasoning job to Gemini 3.1 Pro or a cheap agentic loop to Muse Spark 1.1 still means leaving the seat. Teams that treat this as pure productivity will love the leverage. Teams that treat the seat as architecture should keep data portable and a second surface live for anything mission-critical.
The next model release will still move the Elo board. The more consequential question for the rest of 2026 is which work surface accumulates the habits, connectors, and unfinished multi-hour jobs that are expensive to migrate. Model scores remain the right way to pick a brain. The seat is how that brain stays in the room.