Model Half-Life

From late 2022 to the present, we collected the release dates of headline models from OpenAI, Anthropic, Google, xAI, Meta, Mistral, and major Chinese labs, splitting each vendor by actual product lines to compare release intervals.

The prediction line was calculated using a simple method: adding the median of the most recent 3 intervals for each series to the next release date. Series that appeared only once have no prediction, series that appeared twice use only a single interval, and the predictions weight recent history to reflect current cadence.

The conclusion is clear. 'Model half-life' is less a rigorous concept than a buzzword pointing to the phenomenon of accelerating release speed. Releases are indeed increasing, but it's not a pattern where the release cycle halves every 6 months, and long-term predictions remain weak without more accumulated data.

• The data was built from vendor announcements and release notes, and the author is manually verifying it.
• The classification criteria are sub-series that actually move independently, like Claude Opus vs Sonnet, GPT vs o-series, and Gemini Pro vs Flash.
• Some predictions fall in the past, so certain series are already shown as having a delayed next drop.
• The author has published /model-drops.tsv and is accepting error reports.