Abstract
A machine-readable catalog of large language model API prices covering 76 models across 12 providers, snapshotted 30 August 2026. Published work that estimates the cost of running a language model becomes unreproducible the moment prices move, because the prices it assumed are rarely recorded and vendors change them without notice. This deposit exists so that a cost estimate can cite the dated price basis it actually relied on. Every figure records which upstream source supplied it and the date it was last confirmed. Token-counting methods are labelled explicitly where they are estimates rather than a vendor's own tokenizer, with the assumed characters-per-token ratio stated. Introductory pricing is kept distinct from list pricing rather than silently conflated. The catalog is regenerated daily by a pipeline that reconciles two independent upstream sources against hand-verified vendor overrides. Where the sources disagree, typically reseller pricing against vendor list pricing, the disagreement is flagged for human review rather than resolved silently, and every change lands as a reviewable commit, so the history of any single price can be reconstructed. Included: the catalog as JSON, a flattened CSV for spreadsheet and analysis use, and an OpenAPI 3.1 description of the public keyless API that serves it. Limitations are stated in the included README: prices are as published rather than as billed, token counts for non-OpenAI families are estimates, and coverage is not exhaustive.
Showing the abstract — retrieve the full paper via the Exa API.