Three Reproducible AI Cost Scenarios: Support, RAG, and Agents

Cost advice is only useful when the assumptions are visible. This article publishes three simulated workloads — customer support, retrieval-augmented generation, and a multi-step agent — with enough information to reproduce every result. These are not customer results and they are not claims about typical savings. They are worked examples showing how request volume, fresh input, cached input, output, and agent steps change the arithmetic.

All three scenarios use the GPT-4o mini reference rates stored by TokenCostAI: USD 0.15 per million fresh input tokens, USD 0.075 per million cached input tokens, and USD 0.60 per million output tokens. The dataset was manually reviewed on 2026-08-13 against the OpenAI pricing page. Provider prices and cache eligibility can change, so verify the current official terms before using the results for procurement.

The calculation used in every scenario

Monthly cost = ((fresh input tokens × input rate) + (cached input tokens × cached-input rate) + (output tokens × output rate)) ÷ 1,000,000. Token counts are monthly totals. For agent scenarios, monthly calls = agent runs × average steps per run.

A percentage change below describes only the two stated simulations. It is not a promised saving for another application.

Scenario 1: customer-support drafting

Baseline assumptions: 100,000 requests per month; 1,500 fresh input tokens and 300 output tokens per request; no cached input. Monthly input is 150 million tokens and output is 30 million tokens. The calculation is (150 × USD 0.15) + (30 × USD 0.60) = USD 40.50 per month.

Alternative assumptions: the same 100,000 requests; 300 fresh input tokens, 600 cache-eligible input tokens, and 150 output tokens per request. Monthly cost is (30 × USD 0.15) + (60 × USD 0.075) + (15 × USD 0.60) = USD 18.00. The difference between these two simulations is USD 22.50, or 55.6%. Real cache hits, response quality, and support-message length must be measured before adopting this design.

Scenario 2: reducing RAG context

Baseline assumptions: 50,000 questions per month; 6,000 fresh input tokens and 400 output tokens per question. Monthly cost is (300 × USD 0.15) + (20 × USD 0.60) = USD 57.00.

Alternative assumptions: retrieval and reranking reduce fresh input to 2,500 tokens while output remains 400 tokens. Monthly cost is (125 × USD 0.15) + (20 × USD 0.60) = USD 30.75. The difference is USD 26.25, or 46.1%. This example does not include embedding, vector storage, or reranker charges; those costs must be added for a complete architecture comparison.

Scenario 3: limiting an agent loop

Baseline assumptions: 20,000 agent runs per month; eight model steps per run; 2,500 fresh input tokens and 300 output tokens per step. That produces 160,000 model calls, 400 million input tokens, and 48 million output tokens. Monthly model cost is (400 × USD 0.15) + (48 × USD 0.60) = USD 88.80.

Alternative assumptions: four steps per run after adding a step limit and compact state summary; 1,500 fresh input tokens and 200 output tokens per step. That produces 80,000 calls, 120 million input tokens, and 16 million output tokens. Monthly model cost is (120 × USD 0.15) + (16 × USD 0.60) = USD 27.60. The difference is USD 61.20, or 68.9%. This comparison assumes task quality remains acceptable, which must be tested rather than inferred from cost alone.

Copyable scenario inputs

Support baseline: requests=100000, input=1500, cached=0, output=300. Support alternative: requests=100000, input=900, cached=600, output=150. RAG baseline: requests=50000, input=6000, cached=0, output=400. RAG alternative: requests=50000, input=2500, cached=0, output=400. Agent baseline: calls=160000, input=2500, cached=0, output=300. Agent alternative: calls=80000, input=1500, cached=0, output=200.

How to use these examples responsibly

• Replace the example request counts with measured or explicitly forecast volume.
• Use provider-reported token counts when available; character-based estimates are only planning approximations.
• Measure cache hit rate and confirm which tokens qualify for the provider cache tier.
• Add non-token charges such as embeddings, storage, reranking, observability, and retries.
• Compare answer quality and task completion rate alongside cost.

You can reproduce the arithmetic with the TokenCostAI calculator and review how the reference dataset is maintained on the methodology page. If your result differs, send the exact inputs through the contact page so the calculation can be checked.