Loading & joining 403,495 rows across 3 CSVs…
fin-telligent series
AI Model Shopping Cart
AI Unit Economics: Budgeting Token Spend
When a company adopts AI, the cost model spans 450+ LLMs: model mix to unit price, unit price to budget, budget to P&L.
Raw Table Previews
Joined Table Construction (SQL logic)
Joined Table Preview
520 rows × 22 columns. Every interactive section below is powered by this table.
Source: Artificial Analysis, via Kaggle (LLM API Pricing and Performance Benchmark Tracker)
8 benchmarks, 5 skills.
📡 Source: Artificial Analysis, via Kaggle.
Judged on two dimensions: benchmark score and time to first token (TTFT). ⚡ = fastest third on TTFT.
🧩 No all round champion. Each department picks what it needs.
🏢 Meet our 100-person company. The cart will use these numbers later.
🏢 Every department needs a different mix.
★★★ core ★★☆ important ★☆☆ helpful
💲 Price Tag: what models cost and how 1,000 requests get billed
Three levers set the bill: which tier you pick, whether you cache, whether you batch.
Prices: LLM Price-Performance Tracker · live API rates as collected at snapshot
Tiers are cut on blended price, weighted 3 input tokens to 1 output token. Each tier starts roughly 6 to 10x above the last.
Tier cutoffs observed from the data.
🧮 How the Bill Adds Up
Cost = Input Tokens × Input Price + Output Tokens × Output Price
🔁 Cache · Reuse what you already sent. Repeated input costs 50% less.
Triggers when the prompt starts with content you sent recently. Why cheaper: repeated content needs no recompute.
📦 Batch · Trade speed for price. The whole bill costs 50% less.
Triggers when you submit jobs in bulk and wait up to 24h. Why cheaper: waiting lets providers fill idle GPU time.
Example: Finance / Tax asks to summarize a policy doc · 10K tokens in (8K repeated doc + 2K new question) · 1K tokens out · × 1,000 requests
🗣️
Ask
Input
Repeated tokens + New tokens = Input tokens
🔁
Cached Input
🆕
Fresh Input
💬
Answer
Output
⚡
Real-time
📦
Batch
🧾 Bill
No levers
With cache
Cache + batch
Blended Cost per 1M tokens
Blended = Bill ÷ Total tokens
Unit Budget = (I/O Ratio × Input price + Thinking Multiplier × Output price) ÷ (I/O Ratio + 1)
I/O Ratio = input : output tokens; Thinking Multiplier = extra output for reasoning (1 = no reasoning)
Industry default: I/O Ratio = 3, Thinking Multiplier = 1 brings Unit Budget back to Blended.
Assumption: reasoning models are billed at 2× output (Thinking Multiplier = 2).
Different tasks, different mix.
Same price tag, bigger bill. Reasoning modes think in output tokens.
Illustrative token mix. Cache and batch discounts assumed to stack.
Total Model Cost = Price × Tokens
| Task | Best Estimate of Tokens per Person per Day | How It Adds Up |
|---|
Daily usage is an assumption by task type, adjustable.
🛒 Cart
Pick a department, pick a pack, watch the register run.
Who’s shopping
Tap a department to drop it in the cart. Tap its card’s × to lift it back out.
Pick a Pack
Monthly Cost = Σ ( Team Size × Task Share × Daily Tokens per Person × Days × Unit Budget ) × Batch factor
Daily Tokens per Person = Σ ( Tier Share × Tier Tokens )
Unit Budget = (I/O Ratio × Input price × Cache factor + Thinking Multiplier × Output price) ÷ (I/O Ratio + 1)
Cache factor and Batch factor are 1 when the toggle is off, and 0.6 and 0.5 when it is on.
Task presets are illustrative usage assumptions. 1 token ≈ 0.75 words; 1 page ≈ 500 words.
Reconciled three ways
0-Hallucination Financial Modeling · Finance-Approved Logic · Agent-Scalable
Latest Models & Prices
🔍 Artificial Analysis via Nimble
Profit and Loss Statement (P&L) · Monthly
Fill the Cart to see tokens land here.
