What's New

Amazon Bedrock has launched Advanced Prompt Optimization, a new tool that automates the iterative process of refining prompts for any model available on Bedrock. It combines prompt rewriting with built-in evaluation, allowing customers to compare optimized prompts against original ones across up to five models simultaneously. The tool addresses both model migration scenarios and general performance improvement use cases within a single unified workflow.

How It Works

• Prompt Template Input: Users provide prompt templates along with example variable values (sample user inputs) that represent realistic usage patterns for the prompt.
• Optional Ground Truth: Customers can supply optional ground truth answers to anchor evaluations against known-correct outputs, improving scoring precision.
• Evaluation Criteria: Users specify either a standard evaluation metric or a short natural language description of what "good" looks like, giving the optimizer a clear optimization target.
• Feedback Loop Optimization: The optimizer iteratively rewrites the prompt and evaluates model responses in a closed feedback loop, steering both the prompt structure and wording toward the defined metric.
• Multi-Model Comparison: Up to five models can be evaluated simultaneously — one baseline (the current model) and up to four candidate models — enabling side-by-side migration assessment.
• Multimodal Support: The tool accepts multimodal inputs including JPG, PNG, and PDF files, making it applicable to vision and document-based prompts, not just text.
• Rich Output: Results include the original and optimized prompt templates, evaluation scores, estimated cost per inference, and latency metrics for each model tested.
• Access Methods: The tool is accessible via the Amazon Bedrock Console for interactive use or programmatically through the Bedrock APIs for Advanced Prompt Optimizer.

Why It's Important

• Eliminates Manual Iteration Cycles: Prompt optimization traditionally takes days to weeks of manual trial-and-error; this tool compresses that cycle into an automated, reproducible process.
• Reduces Regression Risk: By evaluating optimized prompts against the original baseline with quantitative scores, teams can detect performance regressions before deploying changes to production.
• Lowers Model Migration Friction: Switching foundation models often requires significant prompt rework; the simultaneous multi-model comparison makes it practical to evaluate migration candidates objectively and quickly.
• Democratizes Prompt Engineering: Teams without deep prompt engineering expertise can leverage automated optimization guided by simple natural language criteria, broadening access across organizations.
• Cost and Latency Transparency: Surfacing cost estimates and latency alongside quality scores enables teams to make informed trade-off decisions rather than optimizing for quality alone.
• Multimodal Readiness: Support for image and PDF inputs means the tool is relevant for the growing class of multimodal enterprise applications, not just text-only LLM use cases.

How It's Different

• Integrated Evaluation Loop: Unlike standalone prompt editors or external optimization libraries, this tool embeds evaluation directly into the optimization loop rather than treating it as a separate post-hoc step.
• Native Multi-Model Benchmarking: Most prompt tools optimize for a single model; Advanced Prompt Optimization natively compares up to five Bedrock models in one run, making it uniquely suited for migration decisions.
• Natural Language Criteria: Users are not forced to choose from a fixed set of metrics — they can describe quality in plain language, which is more flexible than rigid rubric-based evaluators.
• Cost and Latency Output: Few prompt optimization tools surface infrastructure cost estimates alongside quality scores; this tool treats cost and latency as first-class outputs alongside accuracy.
• Fully Managed on Bedrock: Because it runs natively within Amazon Bedrock, there is no infrastructure to provision, no external toolchain to integrate, and access control is handled through existing AWS IAM policies.
• Multimodal Input Compatibility: Support for JPG, PNG, and PDF inputs differentiates it from text-only prompt optimizers common in the ecosystem.

When to Prefer It

• Model Migration Projects: When your team is evaluating whether to move from one foundation model to another and needs objective, side-by-side quality, cost, and latency comparisons before committing.
• Performance Improvement on Existing Models: When you are not changing models but want to systematically improve prompt quality and need before/after evidence to justify the change.
• Regression Prevention During Prompt Updates: When iterating on prompts for a production application and you need assurance that changes improve targeted tasks without degrading others.
• Multimodal Application Development: When building applications that process images or PDFs and need optimized prompts that account for the nuances of multimodal model behavior.
• Teams Without Deep Prompt Engineering Expertise: When the team building the application lacks specialized prompt engineering skills and needs an automated, guided approach to reach acceptable performance.
• Cost-Sensitive Workloads: When you need to balance response quality against inference cost and latency, and want data-driven guidance rather than intuition to make that trade-off.
• Rapid Prototyping and Benchmarking: When you need to quickly evaluate multiple model options for a new use case and want structured output to present to stakeholders or inform architecture decisions.

Availability

• Launch Status: Generally Available (GA) as of May 14, 2026.
• Region Availability: Specific supported regions are not enumerated in the announcement; refer to the official AWS documentation for the current region availability list.
• Pricing: Pricing details are not specified in the announcement; refer to the Amazon Bedrock pricing page for current rates, which likely reflect the underlying model inference costs plus any optimizer-specific charges.
• Access Methods: Available via the Amazon Bedrock Console and the Bedrock APIs for Advanced Prompt Optimizer.
• Model Scope: Compatible with any model available on Amazon Bedrock; up to five models can be compared in a single optimization run.
• Input Limitations: Supports text prompt templates with variable placeholders, plus multimodal inputs (JPG, PNG, PDF); specific token or file size limits are not detailed in the announcement.