--- license: apache-2.0 language: - en tags: - pricing - inference - transparency pretty_name: Thirdeye model catalogue and price sheet --- # Thirdeye model catalogue and price sheet A dated snapshot of what the Thirdeye AI Platform serves and what each token costs, taken from the production API rather than from a rate card held in someone's head. It exists so any price we publish can be checked against a number that was read from a live endpoint on a stated date. ## Files | File | What it is | | --- | --- | | `models_catalogue_2026-11-18.csv` | One row per catalogue model: our price per million tokens, the upstream vendor, and whether the model actually answers on platform credit. | | `models_api_snapshot_2026-11-18.json` | The unedited response body from `GET /api/v1/models`, kept so the CSV can be re-derived and disputed without trusting this README. | ## How this was produced - **Source:** `GET https://thirdeye-ai-alpha.vercel.app/api/v1/models`, an OpenAI-compatible catalogue endpoint, called 2026-11-18 03:50 UTC. - **Prices are ours, not the upstream's.** `prompt` and `completion` are the platform's published rate per 1,000,000 tokens, and they include the platform's margin. They are not what the upstream provider charges us. - **The billed figure is per call, not per model.** Every response carries `x-thirdeye-cost-usd` with the actual charge for that request. That header, not this table, is the authoritative number for anything a customer is told they paid. - **`servable_on_platform_credit` was measured, not assumed.** A one-token chat request was sent to a representative model from each upstream family on 2026-11-18. `200` means the platform serves that model on its own provider keys. `402` with `No API key available for provider` means the row is BYOK only: it appears in the catalogue, but a caller must supply their own provider key, and only metering is charged. ## What this snapshot says, plainly Twelve models are catalogued. Five were probed on 2026-11-18: `gpt-oss-20b` and `ministral-3b` answered on platform credit; `gpt-4o-mini`, `claude-haiku-4` and `deepseek-v3` returned 402 and are BYOK only. The remaining seven rows were not probed and are marked `not probed` rather than guessed, because a catalogue row is not evidence that a model is servable. That distinction is the point of the dataset. A catalogue that lists twelve models while five work is not a lie exactly, but it is a claim that has not been tested, and this file is here to keep the two apart. ## Limits - A snapshot, not a feed. Prices and availability change; the date is in the filename on purpose. - No latency, uptime or quality figures. Those belong to a status page produced by the production path, and inventing them here would defeat the purpose of the file. - Free-tier upstream quotas (for example a 30 RPM ceiling) are not represented. A model can be servable and still rate-limit under load. - Nothing here is investment advice or a statement about the THIRDEYE token. ## License Apache-2.0 for the contents of this dataset.