Problem#
Two OpenAI rate cards have the wrong price. Both are docs-derived and have no entry in discrepancies.
| Model | Our card (per 1M) | Docs standard price | Cause |
|---|
o3 | $1 in / $0.25 cached / $4 out | $2 / $0.50 / $8 | o3.md has two ### Text tokens tables. The first is the standard price. The second is half price (batch or flex). The card uses the second. |
gpt-4o-2024-05-13 | $2.50 / $1.25 / $10 | $5 / $15 (models.dev) | The card comes from gpt-4o.md, which gives the price of the default snapshot gpt-4o-2024-08-06. The older snapshot costs more. |
Source for o3: https://developers.openai.com/api/docs/models/o3.md, section Pricing, read on 2026-10-03.
Impact#
A consumer that uses the card for cost estimates reports half of the real o3 cost.
Ask#
- When a model page has more than one
Text tokens table, use the first (standard) table for base. Put the other table in its own tier, or drop it.
- Do not copy the family page price onto a dated snapshot when the snapshot is not the default one. Leave the card null if no native source gives the snapshot price.
- Check the other OpenAI pages for the same two-table layout.
Acceptance#
GET /v1/models/openai/o3 returns input_tokens: 0.000002, cache_read_tokens: 5e-7, output_tokens: 0.000008 in base.
GET /v1/models/openai/gpt-4o-2024-05-13 returns the snapshot's own price or pricing: null.
- A test covers a docs page with two
Text tokens tables.