GAP-06 · Agent architecture
Prompt Cost and Token Simulator
Current decision
$244 estimated monthly run cost
- Annual estimate
- $2,930
- Cost per call
- $0.0016
- Savings vs same model without levers
- 2%
Same instrument · 3 industries choosing a scenario replaces its inputs; save or share your current changes first
Prefer to read? The two minute case study · problem → approach → metric → outcome
Problem
Before committing to an architecture, leaders need to know whether the workflow can operate within acceptable cost boundaries. A design that works for a pilot may become uneconomic once volume, prompt size, or frontier model share increases.
Approach
The simulator converts call structure into estimated annual cost. It shows the effect of model choice, prompt size, volume, caching, batching, and savings levers.
Why this way
This connects architecture design to run cost, budget exposure, margin, pricing, and financial approval.
The metric
Cost per call and monthly run rate; the monthly delta of switching models.
The trade-off
The cheapest model is not always adequate; caching adds engineering for a real saving.
Outcome
A defensible build versus buy number before anyone draws an architecture box.
Compare with a pinned workload
Baseline: Claude Haiku 4.5, 5,000 calls/day, $244/month. Current change: 0.00 USD/month.
This comparison captures all changed assumptions; it does not attribute a model/volume change solely to caching.
Model
Prompt
≈ 76 input tokensRough estimate (~4 chars/token). Paste a real prompt to size it.
76 in · 400 out
Claude Haiku 4.5
150,000 calls/mo
Illustrative rates · 2026-07-02
Annual cost: billed input + billed output + savings = list-price baseline
Dollar labels round independently. The exact cost bridge below retains the calculated values.
This workload priced across models · monthly, cheapest first
Switching to Google · economy saves $225/mo on this workload, weigh against answer quality for your job.
Savings ladder · monthly, cumulative leverage
Caching + batching cut 2%, $59 / year
Read the exact cost bridge and rates
| Stage | Monthly cost | Reduction from preceding stage |
|---|---|---|
| List price | $249.12 | Baseline |
| + prompt cache | $244.1952 | $4.9248 |
| + batch | $244.1952 | $0.00 |
| Model | Input USD/1M | Output USD/1M | Cached input USD/1M |
|---|---|---|---|
| Claude Sonnet 5 | 3 | 15 | 0.3 |
| Claude Opus 4.8 | 15 | 75 | 1.5 |
| Claude Haiku 4.5 | 0.8 | 4 | 0.08 |
| OpenAI · frontier | 2.5 | 10 | 1.25 |
| OpenAI · economy | 0.15 | 0.6 | 0.075 |
| Google · frontier | 1.25 | 5 | Not supplied |
| Google · economy | 0.075 | 0.3 | Not supplied |
Caching changes the eligible input component first. Batching discounts the eligible share of the remaining total. Savings percentages share the list-price baseline and must not be added.
If you act on this · the call → expected lift → how you'd measure it
The call
Use unit economics to constrain architecture decisions before scaling.
Expected lift · illustrative
Prevents designs that are technically viable but financially fragile.
How you'd measure it
Cost per task, annual run rate, cache hit rate, batching savings, cost per successful outcome.
Steering committee takeaway: Size the call before arguing the architecture. Unit economics often settles the design debate earlier than a diagram does.
How this is built
Stack: Next.js (static) + shared design system; pure client side arithmetic.
Pricing lives in a dated config (`@labs/kit`, as of 2026-07-02), never in copy, with a per-model cache-read price. Tokens are estimated at ~4 chars/token.
Cost/call = input tokens × input price + output tokens × output price. Caching reprices the cacheable share of input at the cache read rate; batching applies a 50% discount to the eligible share.
Limitations: this artifact uses modeled pricing and assumptions. Production forecasting would require current vendor pricing, actual traffic patterns, utilization data, and finance approved costing rules.