Know your models.

The single biggest split — and the one the calculator runs on.

Models cluster into bands: frontier (hardest judgment), strong (high quality at a fraction of the price), efficient (great value), and budget (cheap, adequate). The tier — not the brand name — is what decides whether you're overpaying. A budget model on a task a budget model handles is not 'worse'; it's right.

Two numbers, not one. Output usually costs ~5× input.

Every model has an input price and an output price per million tokens, and output is typically far dearer. That asymmetry is the whole basis of routing: spend frontier dollars where you output little (planning) and wholesale dollars where you output a lot (execution).

Two models at the same price can cost 2× apart.

Price per token is only half the cost — the other half is how many tokens a model spends to finish the job. A terse frontier model can beat a 'cheap' but verbose one on the final bill, because total cost = price × tokens-to-answer. It's the axis that quietly makes 'open source is cheaper' untrue for some heavy open-weight models.

Same quality can arrive 7× faster — or cost 7× less to serve.

Two models with identical answers can differ wildly in how fast and cheaply they're served. That's an infrastructure property (batching, caching, speculative decoding), not an intelligence one — and it's where a chip-constrained lab like DeepSeek just lapped the field.

How much it can hold at once — now often 1M+ tokens.

A bigger window lets a model reason over a whole codebase or a stack of documents in one shot. It also means every new token checks against more history — so long contexts are slower and pricier. Big windows unlock new tasks; they don't come free.

Why two 'big' models can cost 100× apart.

A 'mixture of experts' model can be enormous in total but only activate a slice per request — e.g. a 1.6-trillion-parameter model that fires just ~49 billion per answer. That's the trick behind cheap frontier-ish quality: you pay for what activates, not the full size. It's why headline parameter counts tell you almost nothing about price.

Code and math ≠ creative ≠ judgment.

Models diverge by task. Some draft code and math with near-forced 'right' answers (cheap models tie frontier here); open-ended or high-judgment work is where the frontier still separates. 'Which model?' is really 'which task?' — decide → frontier; execute → cheap.