Amazon Bedrock Introduces Advanced Prompt Optimization for Model Migration and Prompt Tuning

Amazon Bedrock introduces Advanced Prompt Optimization, a new tool designed to automate the migration and optimization of prompts across multiple models. The problem addressed is the manual effort required to re-tune prompts when moving to a newer, faster, or cheaper model—a process that can take days or weeks. The tool uses a reinforcement learning-style feedback loop: it evaluates responses against a user-defined metric, rewrites the prompt, and iterates until optimal. It works with up to five models in a single job and supports multimodal inputs (images, PDFs). Three evaluation modes are available: a custom AWS Lambda function for concrete metrics, an LLM-as-a-Judge for open-ended tasks with a rubric, and steering criteria for brand voice or safety constraints. Users can either migrate from a baseline model to up to four candidates, or improve the prompt on the same model. Outputs include optimized prompts, evaluation scores, time-to-first-token (TTFT), and cost estimates. The tool is accessible via the AWS console, API, or SDK (boto3). Per-job limits: 10 prompt templates, 100 evaluation samples per template, 5 models, 20 text variables per template, 100 multimodal files per sample, and 5 steering criteria per template. Available judge models include Claude Opus 4, Claude Sonnet 4.5, and Claude Sonnet 4.6.