New AI Trick Cuts Data Costs by 99% — And Shows Its Work

Every time you ask an AI to look at a long document, a spreadsheet, or a patient's history, the company pays by the token — the small chunks of text the model reads and writes. Long, detailed records burn through tokens fast, which is why using AI on medical files or security logs gets expensive quickly. A team of researchers has now published TEFM, a method that compresses those long records into compact shortcuts they call "Behavioral Code," keeping the meaning while throwing away the bulk.

The numbers are striking. In tests on clinical data, the system needed only about 1% of the tokens it would normally use. On security data, roughly 2%. That means a job that once cost a dollar might cost a penny, without the AI getting noticeably worse at sorting, flagging, or predicting. The team tested it on three widely used AI models — Qwen3, Gemma-2, and Phi-4 — and got comparable accuracy to standard approaches in every case.

The second half of the breakthrough is trust. Most AI systems give you an answer with no real explanation, which is a problem when the answer affects a diagnosis or a fraud case. TEFM is built to point back to the smallest set of details in the original data that actually drove the decision — like a doctor circling the two lab results that mattered, instead of waving at a 200-page chart.

The catch: this is a research paper, not a product you can buy. It works on structured, table-like data — not free-form essays or chat. And a 99% saving in a lab setting still has to survive messy real-world hospital and bank systems before your bills change.