A real step up from Mercury 2.5, with cheaper tokens than Haiku 5.5 but about the same bill per problem. It was also the only model I tested that found the benchmark answers online.
A post-mortem of the coding agent I built around Mercury 2.5: what moving work into the harness achieved, where the model’s ceiling held, and what I would do differently.
What Mercury 2.5, a fast and cheap diffusion language model, did inside a coding agent I built around it: where it held up, where it failed, and what it cost.
A plain-language summary of Towbin and Plank (2026), Decompression Melting of Peridotite Xenoliths, Journal of Petrology
Notes on Farfan et al. (2025), A Study of the Winston Red: The Smithsonian’s New Fancy Red Diamond, Gems & Gemology
A summary of Shi, Towbin et al. (2024), PyIRoGlass: An open-source, Bayesian MCMC algorithm for fitting baselines to FTIR spectra of basaltic-andesitic glasses, Volcanica
A summary of Moussallam, Towbin, Plank et al. (2024), ND70 series basaltic glass reference materials for volatile element measurement, Geostandards and Geoanalytical Research
A summary of Towbin, Plank, Klein and Hauri (2023), Measuring H2O Concentrations in Olivine by Secondary Ion Mass Spectrometry, American Mineralogist