06-reference

every chatgpt codex merge revolt

2026-07-14·reference·source: Every·by Katie Parrott

"The Urge to Merge (ChatGPT and Codex)" — @Katie Parrott

Why this is in the vault

Two things that matter directly for RDCO: (1) OpenAI's Codex/ChatGPT consolidation is the clearest public signal yet that agentic knowledge work is the contested terrain — both labs are restructuring their flagship products around it, not building separate apps. (2) The Fable-as-orchestrator workflow — expensive model plans, cheaper model executes — is the exact architecture RDCO is implementing, and this issue puts a name and social proof behind it.

Issue contents

SIGNAL — Codex is Dead. Long Live Codex. OpenAI merged its standalone Codex app into a redesigned ChatGPT desktop app, creating three modes: Chat (questions), Work (long agentic assignments across tools), and Codex (developer workflows). The old ChatGPT became "ChatGPT Classic." Power users on X revolted — Theo Browne called it a "generational fumble," Reddit threads documented broken plugins and buried projects. OpenAI's calculus: Codex had 5 million weekly users; ChatGPT has 800 million. The merge bets that agentic features reach more people inside a product they already know, even at the cost of alienating the power-user base that proved the concept.

Anthropic countered the same week by resetting Claude usage limits for all tiers — reportedly to give users more Fable access before it potentially moves to API-only — and unifying Chat and Cowork into a single home tab inside Claude Code. The competitive dynamic Parrott names explicitly: both labs are racing to become the "home for agentic knowledge work" — the interface where any sustained, multi-tool knowledge task begins.

STEAL THIS WORKFLOW — Put Fancy Models in Charge of Affordable Ones The published (non-paywalled) portion covers the Fable-orchestrates-cheaper-models pattern. ChatPRD founder Claire Vo: "What Fable does is none of my business" — she uses Fable 5 as a senior consultant that plans and delegates, with cheaper models doing execution. Every CEO Dan Shipper extends the pattern cross-lab: Fable plans, GPT-5.6 Sol executes, sharing the same brief and project files. The paywalled sections (teased but not accessible here) cover the specific setup, the habit that cut an AI agent's error rate by 53%, and why the model wars benefit users.

Product launch note: Every announced Every All Access at $525/year — a first-party bundle tier adding five products + partner discounts (Cursor, Notion, Framer, Anthropic, PostHog). No third-party sponsor.

Mapping against Ray Data Co

The Fable-plans-cheaper-model-executes workflow is the most concrete answer to the open question surfaced in 2026-07-05-every-tale-two-models — "which model when" for RDCO's agent stack. Fable owns orchestration (planning, review, judgment calls); GPT-5.6 Sol or Sonnet-tier handles bounded execution tasks. RDCO's COO agent already runs on this logic informally; this issue makes it a named, social-proof-backed pattern worth codifying explicitly in the stack SOPs.

The "knowledge-work OS" framing matters for phData DSA context: enterprise clients are about to ask whether they should anchor on ChatGPT (800M users, tool consolidation) or Claude Code (power-user depth, Fable orchestration). RDCO's answer — that the orchestration layer matters more than the chat interface, and labs are converging on architecture anyway — is now backed by a readable Every analysis, not just internal conviction.

The OpenAI revolt story is also a product warning: consolidation for breadth reliably alienates the power-user segment that RDCO sells into. When advising enterprise clients on AI tool standardization, the Codex backlash is a concrete case for keeping specialist workflows separate from general-purpose chat surfaces.

Related