Large language model (LLM)

LLMs are built on the transformer architecture and trained with a deceptively simple objective: predict the next token given everything before it. At sufficient scale — billions to trillions of parameters trained on trillions of tokens — this yields broad capabilities in writing, coding, translation, and reasoning that were never explicitly programmed.

A raw pre-trained LLM only continues text. The assistants people actually use are post-trained with instruction tuning and preference optimization so they follow directions, converse, and refuse harmful requests.

For buyers, the key economic fact is that capability and cost both scale with model size and inference-time compute. Provider catalogs therefore span orders of magnitude in price per token, and matching model tier to task difficulty is the core discipline of managing LLM spend.