TokenCast forecasts agent token spend mid-execution, cutting budget use by 21.3%

SoutheastU · hf · 2026-09-29

Southeast University's TokenCast predicts total token consumption of LLM agent tasks, which can vary by over an order of magnitude across runs of the same task.

Learns a composable cost representation per execution segment covering both its own consumption and context growth; composing segments captures re-reading costs of earlier context in later calls
Forecasts refresh as execution unfolds with no extra LLM calls, at 32.8 ms mean cumulative prediction time on SWE-bench Verified
Reduces MAE vs the strongest comparator by 14.5% on average across 96 combinations (4 suites × 6 agent models)
In offline budget-control replay, uses 21.3% fewer tokens than fixed-budget policy at matched completion

Code: github.com/DEFENSE-SEU/TokenCast.

Original post →

TokenCast forecasts agent token spend mid-execution, cutting budget use by 21.3%

More from coding & agent