Certablo
← Knowledge Base

GenAI Cost & Token Optimization

Optimize spend without sacrificing quality: token accounting, context pruning, prompt compression, response limits, model routing, prompt/semantic caching, batching, concurrency, and capacity-mode decisions.

AIP-C01

Visual overview

GENAI OPERATIONSOptimize only after measuring quality, tokens, latency, throughput, errors, and cost together
01MeasureTokens · latency · quality
02DiagnosePrompt · model · retrieval
03OptimizeCache · prune · batch · route
04ValidateGolden set · alarms · SLOs
Cost and latency optimizations are valid only when the quality target still holds. Re-run evaluation after changing model, prompt, context, cache policy, or retrieval.
AWS SERVICE MAPMeasure before optimizing

Bedrock produces inference/token usage, CloudWatch observes the runtime path, and Cost Explorer supplies broader spend analysis for accountable optimization.

Amazon BedrockInference, caching, and token-producing workload
Amazon CloudWatchLatency, errors, and usage telemetry
AWS Cost ExplorerCost and usage analysis
EXAM-RELEVANT MECHANICS

Technical reference

Optimize the cost of successful work. Token, cache, capacity, and routing metrics need a quality or business-outcome denominator.

Input tokensPrompt + context + history

Prune duplication and irrelevant retrieval while preserving evidence needed for the task.

Output tokensGenerated response

Use response limits and task-specific formats so the model does not generate unused content.

Prompt cacheRepeated supported prefix/context

Use cache checkpoints for reusable prompt portions and measure cache reads/writes rather than assuming a hit.

Semantic cacheEquivalent request → prior result

Application-level reuse requires tenant-aware keys, freshness policy, and a similarity threshold validated against false matches.

Model cascadeCheap-first / capability-aware routing

Route tasks to the least expensive model that satisfies quality, with escalation for harder cases.

CapacityOn-demand / provisioned / batch

Match serving mode to traffic variability, utilization, and latency requirements.

Service limits and capabilities can change. Values shown here reflect the current AWS documentation; use the linked official sources below as the source of truth.

Measure cost at the unit the workload consumes

Foundation-model cost commonly scales with input and output token volume, model/provider, serving mode, and optional surrounding services such as retrieval, reranking, storage, logging, or guardrails. Start with per-request accounting: input tokens, output tokens, model/profile, cache behavior, latency, and business outcome. Aggregate by tenant, feature, prompt version, or workflow so an expensive path has an owner and a reason.

A lower-cost request is not optimized if it fails the task and has to be repeated by a user or escalated to a human. Track cost beside quality and completion rate. Cost per successful task, cost per accepted answer, or cost per automated workflow can be more useful than cost per raw invocation.

Reduce context before changing capacity

Every irrelevant token consumes context space and can increase latency/cost. Remove duplicated boilerplate, retrieve only relevant chunks, compress or summarize stable context when doing so preserves required evidence, and bound conversation history. Prompt templates should make repeated system context explicit so prompt caching can help when supported. Limit output length to what the product actually needs instead of allowing unbounded verbosity.

RAG tuning is often a cost optimization. Better filters, reranking, and chunk selection can reduce the number of low-value tokens sent to the generator. But overly aggressive pruning can remove evidence and increase hallucination. Evaluate after every context optimization and maintain a minimum quality threshold.

Routing and caching trade reuse for complexity

Model routing sends simple work to a smaller/faster model and difficult work to a stronger model. The router itself needs evaluation because false 'easy' decisions can degrade quality. A deterministic rule can be preferable when task classes are known; a learned or LLM router can help with semantic classification but adds cost and uncertainty.

Prompt caching reuses supported repeated prompt portions inside Bedrock. Application-level semantic caching is different: the application recognizes that a new request is sufficiently equivalent to a previous one and reuses a result. That can reduce inference cost dramatically for stable knowledge, but it risks stale or tenant-inappropriate answers. Cache keys must include every context dimension that changes the valid answer, and sensitive/user-specific responses need especially careful isolation.

Batch, concurrency, and capacity must follow demand shape

Interactive workloads prioritize latency and may use streaming or on-demand/cross-Region capacity. Offline summarization, embedding, or evaluation can often use batch-oriented processing where supported. Batch work can improve throughput and reduce request orchestration overhead, but it does not satisfy an interactive response SLO. Schedule non-urgent work away from peak interactive demand when the architecture shares constrained downstream resources.

Provisioned capacity can be economical for stable, sustained demand when utilization justifies the commitment, while on-demand capacity fits variable or uncertain traffic. Concurrency limits prevent one workload from exhausting shared quotas. Cost optimization therefore connects architecture, token efficiency, scheduling, caching, and capacity planning rather than being only a pricing-table exercise.

Key takeaways

  1. 01

    Track input/output tokens, model/profile, cache behavior, latency, and business outcome together.

  2. 02

    Context pruning, tighter retrieval, response limits, and prompt reuse reduce work before capacity optimization begins.

  3. 03

    Model routing is a quality-sensitive classifier; evaluate routing errors as well as downstream model quality.

  4. 04

    Prompt caching and semantic result caching are different mechanisms with different correctness risks.

  5. 05

    Choose on-demand, cross-Region, provisioned, or batch patterns from demand shape and SLOs, not from cost alone.

Official AWS sources

Use these primary AWS resources for the source material behind this article and for deeper reference.