AWS reports that its SageMaker AI multi-turn reinforcement learning workflow raised GPT-OSS-20B’s Wordle solve rate, with training tokens costing about $339.
AWS published a hands-on guide on October 23, 2026, showing how to train a model for multi-step tasks using Multi-Turn Reinforcement Learning (MTRL) on Amazon SageMaker AI. The example trains GPT-OSS-20B to play Wordle, but AWS presents the approach as a template for adapting smaller models to specialized business workflows.
In this setup, SageMaker AI manages training while an agent, task environment, and reward function run on Amazon Bedrock AgentCore. The agent makes repeated decisions, receives feedback as it goes, and returns a single reward when the task ends. That outcome-based signal lets the training process evaluate a full sequence of actions rather than rely on labels for every step. AWS says the same agent could instead run on infrastructure such as Amazon EKS, EC2, or AWS Fargate.
AWS reports that GPT-OSS-20B’s solve rate on held-out Wordle words rose from 61% to 84% at its default reasoning effort. At the low reasoning effort used during training, the rate increased from 15% to 69%. The 100-step run took 12 hours and 40 minutes and covered 105,486 rollouts. AWS calculated its training-token cost at $338.50; AgentCore runtime charges were separate, estimated at about $79 for that run.
The guide describes a serverless training setup that uses a LoRA adapter and Group Relative Policy Optimization. AWS lists Amazon Nova 2 Lite, GPT-OSS-20B, Gemma 4 31B (instruction-tuned), and Qwen 3.6 27B among the supported base models, with availability varying by AWS Region. The sample also tracks training in MLflow and evaluates the model under serving settings. AWS says the training data, agent, and resulting model remain in the customer’s AWS account and are not used to train base models.