1 model, with 1 monitored endpoint (1 actively tracked, 0 no longer) across 1 provider. 1 of it shows at least one detected change — the model version is nominally fixed, so the movement is in how it is served.
Timothée Chauvin, Erwan Le Merrer, François Taïani, Gilles Tredan (2026), "Log Probability Tracking of LLM APIs". ICLR 2026. https://arxiv.org/abs/2512.03816
Timothée Chauvin, Clément Lalanne, Erwan Le Merrer, Jean-Michel Loubes, François Taïani, Gilles Tredan (2026), "Token-Efficient Change Detection in LLM APIs". ICML 2026. https://arxiv.org/abs/2602.11083
@inproceedings{ chauvin2026log, title={Log Probability Tracking of {LLM} {API}s}, author={Timothee Chauvin and Erwan Le Merrer and Francois Taiani and Gilles Tredan}, booktitle={The Fourteenth International Conference on Learning Representations}, year={2026}, url={https://openreview.net/forum?id=hFxivbAgVP} } @inproceedings{ chauvin2026tokenefficient, title={Token-Efficient Change Detection in {LLM} {API}s}, author={Timothee Chauvin and Cl{\'e}ment Lalanne and Erwan Le Merrer and Jean-Michel Loubes and Francois Taiani and Gilles Tredan}, booktitle={Forty-third International Conference on Machine Learning}, year={2026}, url={https://openreview.net/forum?id=7cMlZZYZT0} }