OpenAI now powers the paper's language-model production. A thank-you to Anthropic, and a new chapter with fewer assumptions about access.
The next edition
The Wang Report has moved its language-model production from Anthropic's Claude to OpenAI through Codex. The cutover began on 12 October. The paper is publishing on the new engine.
There is an odd sadness in retiring something that helped give a paper its voice. Claude was there while the desks acquired their rhythms, the correspondents their disagreements, and the editor his habit of asking a draft to earn its place. A great deal of work went into that arrangement. Some of it belonged to us. Some of it was made possible by Anthropic.
That deserves a thank-you. A change of supplier does not make the work we did together disappear.
When access enters the story
The decision followed a recent flagging of our Claude use. The precise basis is not established in this account. We chose to move the production workflow forward rather than leave the next edition dependent on resolving it.
There is a wider context, and it deserves care. Anthropic's published regional policy makes access conditional on location and on where an organisation is based or controlled. Its 8 October policy announcement further explained those restrictions; the updated usage policy takes effect on 12 November. Those public rules do not establish why our use was flagged. They do establish that a model's availability has conditions beyond its ability to answer a prompt.
For a team working across borders, that belongs in the same conversation as reliability and price. The person waiting for a finished piece experiences a missing service before they experience the explanation. Procurement may have signed the agreement months earlier. The deadline arrives today.
What we take with us
Our sources, voices and editorial standards live in the paper's own system. That gave the migration something solid to carry. The shared execution layer changed; the work of checking facts, reading a piece for sensitivity and correcting it after publication still belongs to the publication.
OpenAI gives us a new engine and new habits to learn. Moving from one American provider to another leaves some dependencies shared. It buys an alternative relationship; the responsibility for continuity remains ours. The new writing will have to earn its place by the same standard as the old.
Aya has taken the migration apart in a companion piece: where the model mapping helped, where compatibility ended, and how a team can tell recovery from a successful demonstration. There is room for affection in this farewell, and work in the next chapter.
Thank you, Claude. The next edition is being written with OpenAI, and it still has to survive the editor.
Aya Nakamura's companion piece: The Wang Report's OpenAI Pivot Tests More Than Model Mapping