LLM

An LLM is a model trained on vast amounts of text to predict the word that comes next in a passage. Stacking those predictions one word at a time produces answers to questions, summaries, and translations. According to Anthropic's glossary, pretraining that teaches a model to guess the next word from the preceding context does not by itself produce a model that follows instructions well. Fine-tuning (additional training) afterwards shapes it into a model that answers questions.

An LLM has two stages. Training teaches the model how words fit together from vast amounts of text, and inference runs the finished model to produce answers. Apps and APIs use the inference stage, and API costs are also set at this stage by how much text the model reads and writes.

ChatGPT (OpenAI), Claude (Anthropic), and Gemini (Google) are the names of apps and services that use LLMs. The name stays the same, but the model running inside is replaced whenever a new version arrives. Anthropic sets an end date for older models and guides users to move to a successor model. The Gemini API from Google also has model names that are switched to point at each new release.

That is why the answer to "which service is smartest" changes over time. When comparing, the reliable approach is to look at the name and ratings of the model at that point in time.

No single number decides whether an LLM is good. Ratings in which people compare answers and pick one, and benchmarks that set common problems, measure different abilities. The pages for choosing by use pick one rating to rank by for each use and list models alongside their prices and how much they can read at once. Deciding on your use first and comparing models on the rating for that use leaves less room for doubt.