The early Eocene (49–55 million years ago) is a time interval characterized by elevated surface temperatures and atmospheric CO2 (refs 1,2), and a flatter-than-present latitudinal surface temperature gradient3, 4. The multi-proxy-derived flat temperature gradient has been a challenge to reproduce in model simulations5, 6, 7, especially the subtropical warmth at the high-latitude surface oceans4, 8, inferred from the archaeal lipid-based palaeothermometry, . Here we revisit the interpretation by analysing a global collection of multi-proxy temperature estimates from sediment cores spanning millennia to millions of years. Comparing the variability between proxy types, we demonstrate that the present interpretation9 overestimates the magnitude of past climate changes on all timescales. We attribute this to an inappropriate calibration, which reflects subsurface ocean but is calibrated to the sea surface, where the latitudinal temperature gradient is steeper. Recalibrating the proxy to the temperatures of subsurface ocean, where the signal is probably formed, yields colder -temperatures and latitudinal gradient consistent with standard climate model simulations of the Eocene climate10, invalidating the apparent, extremely warm polar sea surface temperatures. We conclude that there is a need to reinterpret -inferred marine temperature records in the literature, especially for reconstructions of past warm climates that rely heavily on this proxy as reflecting subsurface ocean.