What does it cost?

Why capability is held fixed. Not what models cost — every comparison site answers that, and it answers the wrong question. Prices fall both because models get cheaper and because they get better. This holds capability constant and lets price move.

Both prices are today's. Sellers do not keep last year's rate card online, and the licensed index that does record price history covers only some models — so a genuine then-versus-now is not available for every model, and this does not pretend otherwise. What it compares is the cheapest model that clears the old frontier's level against what the model that set that level still costs — both current, both checkable. The observed series starts the day the stage first ran and accrues forward; the history of fixed levels further up is reconstructed from the licensed index's dated records and is kept as its own series.

The reconstruction prices each past day, it does not observe it. For a date in the past, the bar is the closed frontier as Epoch had it then — Epoch dates every score, so that half needs no price history at all — and both prices are the ones the CC-BY-4.0 AI Price Index records as effective on that date. Nothing about it is mixed into the readings at the top of this section, because the two use different seller sets: the observed series carries seven sellers read first-hand and two registries, the reconstruction carries the licensed index alone. On 14 September the twelve-month bar answered 458× here and 7.6× reconstructed, and the whole of that gap is one disputed registry row the licensed index does not carry. A model also does not enter a reconstructed reading before Epoch dates it, so nothing from 2026 can be the cheapest thing clearing a 2024 bar.

A seller's own dated price wins over the index's copy of it in the reconstruction, exactly as it does everywhere else — but no first-party row we hold can currently be applied to a past date, because a price can only be dated by the day the seller says it took effect and all 456 first-party rows carry no such date. The day we first read a page is when we found out, not when the price changed, and back-projecting it would put today's prices onto two years of history. So the reconstruction is the licensed index throughout, every row records that it applied no first-party number, and the rule will start doing work the moment a seller publishes an effective date.

Blended 3:1 input:output, the field's convention, and held fixed across the whole series. The ratio is stored with every price we record, so changing the convention later cannot silently rewrite this history.

Free and batch tiers are excluded. A rate-limited promotional tier is not a price, and one $0 entry would sink every capability tier to nothing. Batch pricing is roughly half, in exchange for latency measured in hours — a different product, and mixing it in would show a cost halving nobody experiences.

These are the sellers' own published prices. Six labs publish a rate card we can read directly — OpenAI, Anthropic, xAI, Google, Z.ai and Voyage. Where a seller publishes no page we may read, the number comes from a source that can be held to account for it instead: DeepSeek from the CC-BY-4.0 AI Price Index, which carries a dated, confidence-marked record for every rate, and Kimi from two independent MIT registries published only where they agree with each other. The cheapest row below names which seller it came from and, where it was not the seller's own page, which kind of source it was.

How much of this is a seller's own number: 350 read from the seller’s own page · 88 cross-checked between two registries · 63 from the licensed index · 54 from a single registry, marked * · 4 where two registries disagree, marked *. The marked ones are published rather than withheld, and are the best available numbers rather than confirmed ones.

Capability scores are Epoch AI's Capabilities Index, and a model only enters this instrument if it appears in both, matched on name with no fuzzy matching: 118 scored models currently carry a price, and 50 of them clear the tier.

An earlier version of this curve used an aggregator's route price. It has been replaced rather than adjusted. A route price is the price of whichever seller the router happens to use, so it moves for reasons that are nothing to do with a seller's decision — measured on six days of it, no model whose price moved was sold by a single seller and a third of the movements returned to a price the model already held. A rate card does not behave that way, so this is the better input as well as the one we may publish.