AWS shows the full bill for one multi-turn RL run: about $339 took gpt-oss-20b from a 61% to an 84% Wordle solve rate
Following the SageMaker AI training sessions item on October 23, this walks a multi-turn RL run from deploying the reward to evaluation, with the actual spend.
// Key points
- The agent and reward function are hosted on Bedrock AgentCore and invoked by the training job once per rollout; training updates a LoRA adapter with GRPO, and you only define how to score the final result.
- On held-out words the solve rate rose from 61% to 84% at default reasoning effort, and from 15% to 69% at the low effort it trained at.
- The 100-step run used 105,486 rollouts over 12 hours 40 minutes, consuming 491.6 million prefill, 159.7 million sample and 643.3 million train tokens, about $339 at us-east-1 list prices.
- A practical note: at the default sampling_max_tokens of 4096, gpt-oss often ran out of tokens before its first tool call, so it was raised to the 8192 maximum. The post also covers reward-design mistakes.
Builder's takeThe $339 figure is useful: for a multi-step task with clear rules and automatic scoring, one RL run on a small open model now costs less than an outsourced test pass. Rubric-driven follow-up questioning in an AI interview product could try the same recipe, but get the scoring function solid first so the model doesn't learn to game it.
// Background · from #AI agents
Full timeline →- Oct 24 LiteLLM 1.107.0 was a poisoned release: malicious code ran through Python startup hooks and stole credentials without LiteLLM ever being imported
- Oct 23 Anthropic donates the Agent Skills standard to the Linux Foundation's AAIF: more than 70 agent platforms, about 1 million skills published and nearly 280 million installs
- Oct 23 SageMaker AI training sessions are generally available: write your own post-training loop and reward function while AWS runs the GPUs, billed by tokens processed