Satellite-based near-real-time terrestrial monitoring provides timely information on land-surface conditions and changes, supporting both rapid assessment and predictive modelling. Converting satellite observations into timely information remains challenging because these observations are irregular and incomplete, while most algorithms require compositing or fixed temporal sampling, introducing weeks to months of additional latency beyond that imposed by satellite revisit frequency. Here we present Terra-GPT, a phenology-aware generative pre-trained Transformer with >130 million 30 m resolution Landsat and Sentinel-2 observations. Terra-GPT operates directly on arbitrary-length cloud-free time series and reduces algorithm-induced latency to near zero by leveraging encoded phenological dynamics to support retrieval even on dates without satellite observations. Four Terra-GPT downstream applications are demonstrated: daily live fuel moisture content, field-scale soil moisture, within-season crop type mapping, and crop damage detection. Terra-GPT consistently outperformed supervised baselines, existing foundation models such as NASA–IBM Prithvi, and operational products such as the NASA soil moisture product across the demonstrated applications. These results establish phenology-aware time-series foundation models as a scalable paradigm for monitoring rapid land-surface change, moving beyond foundation models designed primarily for static mapping. Terra-GPT provides a unified framework for timely monitoring across wildfire, hydrology and food-security applications. Owing to its phenology-modelling capability, Terra-GPT can be extended to other large-scale biophysical retrieval tasks that support climate-resilient decision-making.