Most organizations running AI on AWS Bedrock have access to the same underlying operational data. Very few have the tools to turn that data into decisions. Model Optimizer is an AI operational intelligence platform built on a single operating principle — Evidence → Intelligence → Decision.
The platform isn't organized around features. It's organized around the operational decisions AI teams need to make.
Model Optimizer is an AI operational intelligence platform built on a single operating principle: Evidence → Intelligence → Decision. Every capability on the platform follows that sequence.
The invocation record is the evidence — primary-source operational data generated by your own AWS Bedrock environment.
The monitoring, capability, fitness, and prompt analysis layers are the intelligence — the insight your team can act on.
The model selection, prompt rewrite, and cost optimization actions are the decisions — the outcomes every surface exists to support.
No step is optional. No step is approximate. Every recommendation traces back to observable operational data from your own AWS environment.
Evidence without intelligence produces data overload. Intelligence without evidence produces assumptions dressed as insight. Decisions without either produce outcomes that cannot be explained, trusted, or repeated.
Model Optimizer: Managing AI through evidence.
Model Optimizer starts where every AI optimization decision should start: primary-source operational data.
AWS Bedrock Model Invocation Logging records every model call made through your Amazon Bedrock environment — request and response payloads, token counts, latency, model identifiers, account information, regions, error states, and inference parameters. It is primary-source operational data generated directly by AWS infrastructure, delivered to your own S3 bucket, and read by Model Optimizer from there.
Your data stays in your environment. Model Optimizer reads from it with read-only access.
That data foundation is what separates Model Optimizer from platforms working from sampled telemetry, API wrappers, or aggregated reporting. The intelligence layer is only as good as the evidence beneath it.
The intelligence layer transforms invocation data into actionable operational insight across four capabilities.
Continuous operational visibility across your entire AI environment. Cost by model, token consumption by prompt pattern, error rates by type and region, usage trends over time, and account and regional breakdowns — all derived from the same invocation data layer.
Visibility is not the goal. It is the prerequisite.
Benchmark-driven evaluation of what every model in your environment can do — across 11 standardized task dimensions.
Model Optimizer evaluates every model using published benchmarks where credible, widely adopted evaluations exist, and proprietary evaluations where they do not. Every score includes a rank, a percentile, and a source attribution. The Capability Radar translates those scores into a single visual profile — the human visual system recognizes patterns faster than it compares numbers, and the shape of that profile makes cross-model capability differences immediately readable.
The evaluation layer that moves from what a model can do to whether it is the right model for what you are actually doing.
Model Fitness anchors capability evaluation to your actual deployment — your current model, your prompt patterns, your workload distribution, and your cost constraints. Every fitness view surfaces rank, percentile, and cost delta for every alternative in the benchmarked field. The value picks filter narrows the field to models that offer the strongest combination of performance and cost efficiency.
Capability tells you what a model can do. Model Fitness tells you whether it is the right model for the work you actually do.
Active improvement of the production prompts running in your AI environment — not sample prompts, not templates, the real prompts identified from your invocation logs.
Model Optimizer analyzes each prompt for issues by severity, scope, and issue type — negative phrasing, missing source grounding, excessive redundancy, contradictory instructions — and produces a rewritten version with every change explained. The Playground lets you test the rewrite against your current model for token efficiency and against one benchmarked alternative model before any change goes into production.
The intelligence capabilities above exist to support one outcome: better AI operational decisions.
The Decision Layer enables decisions such as:
Optimization without monitoring is a hypothesis, not a result.
Evidence becomes intelligence, intelligence drives decisions, and validation loops the whole sequence back to evidence as your workloads evolve.
AWS Bedrock Model Invocation Logging (MIL) — delivered to your S3 bucket; Model Optimizer reads from there.
AI Monitoring · AI Model Capability + Capability Radar · Model Fitness · Prompt Optimization + LLM Benchmarking Methodology
Model Selection · Prompt Rewrites · Routing Recommendations · AI Cost Optimization
Continuous Monitoring · Outcome Confirmation · Next Optimization Cycle
↑ Loops back to Evidence as workloads evolve.
Every page on this site maps to one layer of that sequence. Every capability on the platform is grounded in the same operational data. Every decision traces back to evidence.
Your AWS Bedrock environment is already generating the operational data you need. Model Optimizer connects to it, analyzes it, and surfaces the intelligence required to improve AI capability and reduce AI operating cost — starting with a free account. No code changes. Read-only access via IAM role.
Get Started FreeModel Optimizer is an AI operational intelligence platform built on AWS Bedrock Model Invocation Logging data. It surfaces the intelligence required to reduce AI operating costs, evaluate model capability and fitness, optimize production prompts, and monitor your AI environment continuously. Every capability follows the same sequence: Evidence → Intelligence → Decision.
Model Optimizer reads AWS Bedrock Model Invocation Logging data delivered to your own S3 bucket. It reads from your bucket with read-only access. Your data stays in your AWS environment — Model Optimizer never stores your invocation records.
Most AI monitoring tools work from sampled telemetry, API wrappers, or aggregated reporting. Model Optimizer reads primary-source invocation data at the individual call level — every prompt, every token, every response, every error — from AWS infrastructure itself. That data foundation is what makes the intelligence layer accurate rather than approximate.
No. Model Optimizer reads from your existing AWS Bedrock Model Invocation Logging data. If Model Invocation Logging is already enabled and delivering to S3, setup requires no code changes. If it is not yet enabled, Model Optimizer provides a setup guide that walks through the configuration process.
Model Optimizer requires Amazon Bedrock with Model Invocation Logging enabled and logs delivered to an Amazon S3 bucket. No additional AWS services are required for basic connectivity. CloudFormation is used to create the read-only IAM role that grants Model Optimizer access to your S3 bucket.
Full Analytics Mode reads prompts and responses from your S3 bucket to provide complete optimization intelligence — including prompt analysis, rewrite recommendations, and Playground testing. Organizations connecting in Full Analytics Mode execute a formal signed data handling agreement. Privacy Mode reads metadata only — prompts and responses remain in your AWS account and are never accessed by Model Optimizer.
Most organizations complete the full setup in under 30 minutes. If Model Invocation Logging is already enabled and delivering to S3, setup can be completed in minutes.
Optimizing Your AI walks through the full optimization workflow in applied terms. The Free Bedrock Model Invocation Logging Setup Guide covers the AWS configuration process in detail and is available free with a Model Optimizer account.