Skip to content
Carlos KiK
Go back

OpenAI Is Merging Chat and Work Into One Surface, With One Token Pool

On November 2, 2026, OpenAI sent users an email confirming a quiet product consolidation: the Chat and Work tabs are merging into a single surface. Codex, the coding agent, stays separate. The two surfaces will draw from a shared token allowance pool instead of separate budgets.

The structural point is bigger than the tab layout. For two years the industry has drawn a line between “chat with an assistant” and “send work to an agent”, and priced the two differently. Inside OpenAI the split was literal: two tabs, two mental models, two allowance pools. The merge says out loud what the usage patterns already said: the same person chats at noon and delegates a research task at two, and forcing those into different products with separate meters never matched how the work actually flows.

The shared token pool is the detail with real consequences. When one surface and one meter serve both modes, a heavy agent run visibly competes with everyday chat for the same allowance. That turns subscription value into one number, and it turns every agent workflow into a visible cost against the same budget the user checks daily. Watch for behavior changes here: people who were relaxed about burning chat tokens may think twice when the same pool funds a long agent session.

There is a precedent. Anthropic moved the same direction earlier, unifying chat and Cowork into one home where sessions and files follow the account across desktop, web, and mobile. Two frontier labs converging on the same shape, one interface spanning conversation and delegated work, suggests the split was a transitional design, not a durable category. Codex remaining separate is the interesting counterpoint: OpenAI still believes coding agents deserve their own surface, or at least their own meter, for now.

For builders the takeaway is not about tabs. It is that the “assistant product” and the “agent product” are collapsing into one surface with one budget, and the products you build on top of these subscriptions should assume that same shape: one user, one allowance, mixed modes, and no separate meter to game.


Share this post on:

Previous Post
Nvidia Open-Sourced Auto Ontology, a Framework for the Definitions Agents Depend On
Next Post
Reality Tested: Dyson CameraJet