AI Model Release Tracker

What this tracks

This page tracks the language models you might actually choose between: the current frontier models from the major API providers and the most capable open-weight releases. For each, it records the specifications that decide whether a model fits a job — context window, maximum output, input and output modalities, licence, and how you access it.

It is a living reference, not a news feed. When a provider ships a new model or revises a spec, the change is read from the source, staged, and checked by a person before it appears here.

How to read it

Every value in the table is read from a primary source — a provider's own model documentation, or, for open-weight models, the model's config.json and model card on Hugging Face. Follow any value to the page it came from.

A dash ("—") is not missing data. It means no primary source states that value, so we do not either. Two cases in this first release:

• Release dates appear only where a provider states one. For an open-weight model, a repository's creation date is not the same as an announced release, so the field is left blank rather than imply a date the source does not give.
• Google's Gemini models are absent for now. Their published model pages render the specifications with JavaScript that our sourcing cannot read cleanly, and we will not transcribe numbers we cannot check against the page. They will be added once sourced.

Why the numbers can be trusted

The specifications here are not recalled from memory — recalling a context window is exactly the kind of plausible-but-wrong claim this site exists to avoid. Each value is read from its source on a known date, and a person confirms it against that source before it goes live. The record of who verified a value, and when, is only ever set by that human check.

Context windows deserve one caveat. A "1M token" window does not hold a fixed number of words: the same text produces different token counts under different tokenizers, so a window measured in tokens and the same window measured in words can move independently. Where a provider states a token figure, that figure is what appears here.

Who should care

If you are choosing a model to build on, four fields usually decide it: context window (how much you can put in), maximum output (how much you can get back in one response), modalities (whether it reads images), and licence (whether you can run the weights yourself). This table places those side by side, sourced, so the comparison is the provider's own numbers rather than a summary of them.