What's new in AI

Active: on sale or supported today. Legacy: superseded, or scheduled to shut down; weights of open models stay downloadable. Retired: shut down or never publicly released. 183 models were added automatically from live feeds and are marked New; we have not written notes on those yet. It is not literally every model ever made (thousands of fine-tunes exist), but it covers the major labs and the landmark open models.

Status for Claude, GPT and Gemini models comes from the vendors' own deprecation pages: Anthropic ↗ (opens in a new tab), OpenAI ↗ (opens in a new tab), Google ↗ (opens in a new tab). Those pages change often; a shutdown date in the live feed also marks a model Legacy.

“To run it locally”: memory is what it takes just to load the weights, so add headroom for long contexts. Where the developer or a reliable guide gave a figure we use it and name the format (for example “native FP8”). Otherwise we estimate about 0.6 GB per billion parameters at 4-bit quantization, or 2 GB per billion at 16-bit. Mixture-of-experts models still need every expert in memory, so use the total parameters, not the active ones. These are planning estimates, not guarantees: check the model card before buying hardware. To learn how this works, start with the learning path on how large language models work.

Where the data comes from: OpenRouter model list (reachable), the Arena leaderboard (LMArena, CC BY 4.0) (reachable), Hugging Face (reachable). Refreshed automatically every 30 minutes. Our own notes on running cost, licences and retirements are written by hand and layered on top. A model is listed when it appears in these feeds; that is not an endorsement.