Building products for the intelligence that is coming

When OpenAI released GPT-6 in early September, we began evaluating it in Narrative within days. A new model usually moves our evaluation results by a few percentage points. This one was a step change, and more capable models are expected to follow.

Products can hold intelligence back

It is surprisingly easy to build a product that limits what a better model can do. Rigid prompts can force a new model to follow logic written for an older one. Output formats designed around the limits of earlier models can hide what a newer model is capable of. The product then improves by less than the model does.

Designing for the next model

We try to build Narrative Coach so that new capability flows into the experience as soon as it is available. In practice that means:

• describing the outcome we want from a model and leaving room for it to reason about how to get there;
• keeping the business’s own data and context well organised, so a better model has more to work with;
• running every new model against the same evaluations before it reaches customers, so we know what has improved and what has not.

What this means for banks and lenders

For an institution choosing an AI partner, a useful test is whether the product can take on the next generation of models safely. Two questions help: how quickly can a new model be adopted, and how will you know it is better before your customers see it?