About

Every score says whether the lab reported it or someone else measured it, and the two are never mixed. A change against the previous model is shown only when both scores come from the same test, version and source; otherwise the entry says no change is shown, and why. For each release the record tracks which disclosure documents the lab published, and marks the ones we have not yet checked, so changes in what labs disclose can be followed year by year.

The registry is built and maintained by Walker Robinson. Each night a script checks four sources for new releases, and a language model drafts an entry for each one. Nothing appears until a person approves it. The record reaches back to 2022-11-30, built from Epoch AI's datasets, and the full dataset is downloadable under CC-BY 4.0. The methodology says where AI tools were used. The 325 entries built from Epoch AI's data were generated by script; their disclosure links were approved by Walker Robinson.

The primary record of a model release is the lab's own announcement, and most citations lead to it. The party being measured writes it, and it reports the benchmarks the lab chose to run. It gets edited after the fact, and sometimes it comes down. What a lab did not say leaves no trace. A year on, it is hard to say what shipped and what was disclosed at the time.

Recording each release in one place pins down the date, the source URL, the retrieval date, the benchmark and its version, and which disclosure documents we could find. For releases before July 2026 we checked in mid-2026, and again in October 2026 for 161 releases, so a document published after launch still counts. A claim about how fast models are released can then be checked against a count, and a claim about disclosure against the record, including the entries where the answer was nothing. The dataset is published in full, so anyone can check a conclusion here against the same evidence. The methodology page states the rules and the limits. To report an error, open an issue on GitHub.