Two findings, two different documents

One study cut the instructions and lost nothing. The other let the record build up in detail and gained. They are not in conflict, because they changed different things.

Instructions: cut them

Share of one AI tool's standing instructions removed

Measured on the maker's own coding tests, for one generation of models. Not a public benchmark, and not tested on other kinds of work.

The record: let it build up

Improvement over the strongest existing methods

Achieved with a modest, freely available model, and by the whole method rather than by length alone. Peer reviewed, accepted to ICLR 2026.

Different documents, different systems, different ways of measuring. The two numbers should not be read against each other. The direction of travel is the point, not the size of the gap.

Sources: Anthropic, "The new rules of context engineering for Claude 5 generation models", 24 July 2026. "Agentic Context Engineering", Stanford, SambaNova and UC Berkeley, arXiv:2510.04618, ICLR 2026.