Prompt habits from older AI models now raise cost and can lower quality

Prompts pile up debt the way code does, and upgrading to a newer AI model exposes it. That was the argument of Lance Martin of Anthropic in a talk published on the AI Engineer show on 6 November 2026. Habits teams added to steady older models can now make newer ones costlier and sometimes worse.

What was said

Martin opened his first point with a plain claim: "prompts accumulate debt just like code." He listed habits from earlier model generations. They included telling the model to think step by step, to double-check its work, and to be maximally thorough. Some teams also gave it a scratch pad to think in.

He said newer models no longer need this. In his words, "They don't need to be told to be thoughtful. They're already very thoughtful." He added that these habits raise cost and can lower performance.

He gave a second case. Earlier models were weaker at following instructions, so they would drop one of two rules that contradicted each other. The prompt still worked. Frontier models, he said, often get confused by such rules.

To show the effect, Martin built a synthetic test of about 50 customer support tickets. He simulated a move from one Opus model to a newer one, injecting each old habit in turn. The move alone brought some cost savings and a small performance gain. Running Anthropic's prompt audit command on top gave a further gain in both, because it stripped out habits that made the model spend extra tokens.

Why it matters

Our reading: a model upgrade is not finished when the new model is switched on. In Martin's test, the swap alone captured only part of the gain. The rest needed someone to read the old prompts and remove what no longer belongs.

That work has a price. Someone must review prompts, rerun a test set and decide what to cut. Leaving the rituals in also has a price, because Martin said they add tokens and can add confusion. Managers who plan a model migration may want that review as a line in the plan.

The other side

The evidence here is limited. It is one synthetic test, built by an employee of the company that makes the models. Martin did not claim that every team will see the same savings.

He also said the boundaries are fuzzy. He could not say precisely how far back in model history these habits stop helping, and called it "squishy". A weaker model may still need more prompting, and he said he did not know whether that is true.

Martin also did not say that teams fail to budget for this cleanup. His message was that the fix is simple and worth trying first. The budgeting gap is our inference, not his claim.