A luxury House lives on what it remembers: the cut of a jacket from 1987, the exact shade of a red, a client’s preference for unlined gloves. This memory is its most fragile asset, and until now artificial intelligence has treated it as an expense. Each request meant rereading the entire file, and each rereading was billed.
OpenAI’s announcement of GPT‑6.1 Sol shifts that calculation. The provider lists for its new model 2 dollars per million input tokens and 10 dollars per million output tokens, one fifth of the rates of its flagship model, Astra (10 and 50 dollars). The most interesting figure, however, appears further down in the press release: context already transmitted, known as “cached”, costs only 0.10 dollars per million tokens, 95% less than standard input.
Translated into uses, an agent that keeps in mind a House’s style guide, its collection archives and a client’s history can reopen that file at every exchange for a fraction of the price. What was a laboratory luxury becomes an ordinary budget line.
These figures come from the provider itself, measured on benchmarks it chose. OpenAI acknowledges, moreover, that its factuality tests rely on deliberately difficult questions, not representative of everyday use. On the agent’s transparency when a search tool is down, the model still fails in 2.1% of cases. For a House in which every sentence commits its name, that residue is far from trivial: a cheap agent that invents a founding date costs more than an expensive agent that stays silent.
The question therefore changes in nature. It is no longer about what one can afford to have the machine read, but about what one accepts to entrust to it. When memory costs almost nothing, selection becomes the rare gesture: deciding which archive collections enter the context, which remain out of reach, and who rereads what comes out. A know-how was once passed on through the workshop. From now on, one will have to choose what one accepts to have a third party memorise.
Cette publication est également disponible en :
