Multi-Turn RL on SageMaker AI with AgentCore

Using gpt-oss-20b, solve rates improved from 61% to 84% in default reasoning mode and from 15% to 69% at low effort. Training costs approximately $339 across 105,486 rollouts, with train tokens dominating due to trajectory length. The system uses LoRA adapters and runs on a serverless architecture with full control over base models and training cost by token type.

Summarised from the original article by an open-source model, then checked automatically: every number and English name must appear in the source. The original article is authoritative.