In an AWS Training Run, the Train Tokens Ate the Budget
An agent that files an expense report or triages a support ticket takes a sequence of actions, and each step depends on what came before.
AWS's MTRL service is serverless: you pick a base model and it provisions the training and sampling infrastructure. You pay per million tokens in three dimensions, priced by base model.
In its 100-step run, AWS reports 491.6 million prefill tokens, 159.7 million sample tokens, and 643.3 million train tokens across 105,486 rollouts in 12 hours 40 minutes. At GPT-OSS-20B list prices in US East (N. Virginia) at the time of writing, that comes to about $339. The train dimension dominates and scales with trajectory length, so shorter games and less verbose reasoning cut the bill; in this setup, an agent's verbosity is a line item.
AI-generated discussion · Different perspectives
The line item that should worry agent builders is the train dimension. It is the largest of the three in this run and it grows with the length of the trajectory, so a chatty agent is an expensive agent.
A lower per-token price helps, and AWS chose GPT-OSS-20B partly because its reasoning effort is controllable per request. But the train dimension still scales with trajectory length, so the bigger lever is getting the agent to stop explaining itself.