Reference grid, 6 × 5, 30 figures
Token cost vs cache hit reference grid
The monthly cost of a fixed LLM research loop for every combination of input-token size and prompt-cache hit rate across this grid, priced on Claude Sonnet 4.6. Each cell is a live run of the Token Cost Optimizer engine; no value on this page was entered by hand.
Monthly cost scales linearly with input-token size and falls steeply as the prompt-cache hit rate rises, because cached input reads price far below fresh input tokens. Across this grid the monthly cost runs from $84 (smallest prompt at 2000 tokens, 90% cache) up to $772 (heaviest prompt at 64000 tokens, 0% cache). At the heaviest prompt, lifting the cache hit rate from 0% to 90% cuts the bill by 77% — the single largest lever on this surface. For how to amortize the cache write across a loop, see token-cost cache amortization. Education only — not investment advice.
Monthly cost by input tokens × cache hit rate
Rows = input tokens per call. Columns = prompt-cache hit rate. Each value
is the engine's
costPerMonth output in USD.
| tokens \ cache | 0% | 25% | 50% | 75% | 90% |
|---|---|---|---|---|---|
| 2,000 | $103 | $97 | $92 | $87 | $84 |
| 4,000 | $124 | $114 | $104 | $93 | $87 |
| 8,000 | $167 | $147 | $126 | $106 | $94 |
| 16,000 | $254 | $213 | $172 | $131 | $106 |
| 32,000 | $427 | $345 | $262 | $180 | $131 |
| 64,000 | $772 | $608 | $444 | $280 | $181 |
Headline metric: monthly cost (USD). The CSV download below also carries the cost per idea, cost per validated trade, and effective cost per call.
How the grid is made
- Engine
- Token Cost Optimizer
(
token-cost-optimizer), computed live from/engines/token-cost-optimizer.js - Grid
- input tokens ∈ {2000, 4000, 8000, 16000, 32000, 64000} × cache hit ∈ {0%, 25%, 50%, 75%, 90%} = 30 cells
- Fixed inputs
- model=claude-sonnet-4-6, output_tokens=1500, calls_per_idea=5, retry_rate=0.2, ideas_per_day=30, validation_rate=0.25
- Computed
- , recomputed in CI on every build
The engine is deterministic: token accounting and cache-read pricing are closed-form, so the same input always returns the same output. The full method — the per-call token math, the retry multiplier, and the cache discount — is documented at the Token Cost Optimizer methodology page. For tying spend back to a per-trade unit, see the cost-per-validated-trade framework.
Reference grids are planning aids, not financial, tax, or investment advice. Prices reflect the engine's published rate card and may differ from your billing.