Why this is in the vault
Thompson documents OpenAI's deliberate pivot from chat to agent/code interface — killing ChatGPT-as-chatbot and renaming Codex as the new ChatGPT — with Claude Cowork named as the direct competitive forcing function.
The core argument
OpenAI shipped "ChatGPT Work," combining ChatGPT and Codex into a single superapp powered by GPT-5.6. The shock: the Codex Mac app literally renamed itself ChatGPT, while the old ChatGPT became "ChatGPT Classic." The new app is not a chat app — chat is a sidebar afterthought. Thompson frames this as OpenAI acknowledging that Codex is categorically more capable than ChatGPT (calculator vs. Excel), and making a high-stakes bet: trade the enormous chatbot user base to forcibly onboard non-coders into agentic workflows.
Thompson endorses the product vision personally — he runs Codex headless on a rack-mounted Mac Mini and accesses it from any device, calling it "a glimpse into the future." But he's skeptical of the go-to-market: Codex is powerful and also intimidating; "work" branding may repel consumer use; and the simultaneous push into advertising is in tension with the product being framed as a productivity tool, not a consumer app.
The competitive frame is explicit: OpenAI built ChatGPT Work as a direct answer to Anthropic's Claude Cowork (launched January 2026). The AI coding/agent layer is where the real differentiation lives — and it's territory Apple will never enter, Meta lacks user trust for, and Google appears unable to ship.
Mapping against Ray Data Co
Claude Cowork being the forcing function for OpenAI's biggest product pivot in years is the sharpest signal yet that the Anthropic ecosystem is winning the enterprise agent race — which is the exact stack RDCO operates on at phData. If Thompson's read is right that Codex/agentic interfaces are replacing chat as the primary AI surface, the Claude Code-first workflow already in place at RDCO is ahead of where the market is heading, not behind. Concretely: any phData client still orienting around ChatGPT-as-chatbot for knowledge work is running the product OpenAI just deprecated. That's a positioning wedge — RDCO can frame Claude Cowork/Claude Code adoption as the current-generation stack vs. a client's legacy ChatGPT chat deployment.
The advertising tension Thompson raises is worth tracking: if OpenAI is splitting personality between consumer ad-supported chat (mobile) and pro agentic work (desktop), Claude's enterprise-only lane stays cleaner. No mixed signals on business model.
Related
- [[2026-06-24-stratechery-ben-thompson-vibe-coding-adventure]] — Thompson's personal Codex-on-Mac-Mini experiments; directly foreshadows this pivot
- [[2026-06-15-stratechery-ben-thompson-anthropic-safety-superpower]] — Thompson on Anthropic's competitive moat; safety-as-feature vs. OpenAI's feature-velocity approach
- [[2026-05-09-innermost-loop-singularity-fermi-paradox]] — noted the OpenAI superapp rumor (ChatGPT + Codex + Atlas browser bundled) two months before it shipped
- [[2026-04-25-saas-death-thesis-vault-synthesis]] — synthesizes the thesis that chat replacing UI is the mechanism of SaaS disruption; this article is a live data point