Why the same model has two prices
Ask what a model costs and you will get two different answers, and both can be right. The list rate is what the model's own vendor charges you directly. The gateway rate is what a reseller charges for the same weights, and it moves on its own calendar. Below is every model for which we can cite both.
https://aicomp.ai/v1).
Create one free →
curl https://aicomp.ai/v1/models \
-H "Authorization: Bearer sk-your-gateway-key"A JSON list of model IDs means the key is good. Invalid token means it was copied wrong.
Both rates, side by side
| Model | Vendor | Official list in / out |
Gateway realtime in / out |
Below list in / out |
|---|---|---|---|---|
| claude-3-5-haiku-20241022 | Anthropic | $0.8 / $4 | $0.4 / $2 | 50.0% / 50.0% |
| claude-3-5-sonnet-20241022 | Anthropic | $3 / $15 | $1.5 / $7.5 | 50.0% / 50.0% |
| claude-3-7-sonnet | Anthropic | $3 / $15 | $1.5 / $7.5 | 50.0% / 50.0% |
| claude-3-haiku-20240307 | Anthropic | $0.25 / $1.25 | $0.125 / — | 50.0% / —% |
| claude-3-opus-20240229 | Anthropic | $15 / $75 | $7.5 / — | 50.0% / —% |
| claude-3-sonnet-20240229 | Anthropic | $3 / $15 | $1.5 / — | 50.0% / —% |
| claude-fable-5 | Anthropic | $10 / $50 | $5 / $25 | 50.0% / 50.0% |
| claude-fable-5-1 | Anthropic | $10 / $50 | $5 / $25 | 50.0% / 50.0% |
| claude-haiku-4-5-20251001 | Anthropic | $1 / $5 | $0.5 / $2.5 | 50.0% / 50.0% |
| claude-opus-4-1-20250805 | Anthropic | $15 / $75 | $7.5 / $37.5 | 50.0% / 50.0% |
| claude-opus-4-20250514 | Anthropic | $15 / $75 | $7.5 / $37.5 | 50.0% / 50.0% |
| claude-opus-4-5-20251101 | Anthropic | $5 / $25 | $2.5 / $12.5 | 50.0% / 50.0% |
| claude-opus-4-6 | Anthropic | $5 / $25 | $2.5 / $12.5 | 50.0% / 50.0% |
| claude-opus-4-7 | Anthropic | $5 / $25 | $2.5 / $12.5 | 50.0% / 50.0% |
| claude-opus-4-8 | Anthropic | $5 / $25 | $2.5 / $12.5 | 50.0% / 50.0% |
| claude-opus-5 | Anthropic | $5 / $25 | $2.5 / $12.5 | 50.0% / 50.0% |
| claude-sonnet-4-20250514 | Anthropic | $3 / $15 | $1.5 / $7.5 | 50.0% / 50.0% |
| claude-sonnet-4-5-20250929 | Anthropic | $3 / $15 | $1.5 / $7.5 | 50.0% / 50.0% |
| claude-sonnet-4-6 | Anthropic | $3 / $15 | $1.5 / $7.5 | 50.0% / 50.0% |
| claude-sonnet-5 | Anthropic | $2 / $10 | $1 / $5 | 50.0% / 50.0% |
| gpt-5.6-luna | OpenAI | $0.2 / $1.2 | $0.1 / $0.6 | 50.0% / 50.0% |
| gpt-5.6-terra | OpenAI | $2 / $12 | $1 / $6 | 50.0% / 50.0% |
| gpt-6-astra | OpenAI | $10 / $50 | $5 / $25 | 50.0% / 50.0% |
| gpt-5.6-sol | OpenAI | $4 / $20 | $2.5 / $15 | 37.5% / 25.0% |
23 of these 24 sit at exactly 50% below list on input. The highlighted row is the exception, and it is the reason you should not assume a flat rate.
What the table actually says
Two findings, and the second one is the useful one.
First: the discount is remarkably uniform. 23 of these 24 models sit at exactly 50% below list on input. That consistency is real, not rounding — it is what the upstream promotion looks like when applied across a catalogue.
Second: the one model that breaks the pattern is the most informative row in the table. gpt-5.6-sol sits 37.5% below list on input and 25% below on output, not 50%. The reason is that OpenAI lists its $4/$20 rate as promotional, valid at least through 21 November 2026. A promotional rate on one side of the comparison moves the percentage on both.
This is precisely why this site does not state a flat rate anywhere. A single exception with an expiry date is enough to make any blanket claim false, and the exception is not an obscure model — it is one of the four OpenAI models with a published list price.
Official prices move on their own schedule — and can be cancelled
The list side is not stable either. Anthropic launched Claude Sonnet 5 at $2/$10 per million tokens with a scheduled rise to $3/$15 on 1 September 2026. On 10 August it reversed that: the launch rate became the standard rate and the September rise no longer applies. Sonnet 5 is $2/$10 today, and that is confirmed against Anthropic's own changelog rather than inferred from a launch post.
If you had hardcoded if date >= 2026-09-01: sonnet = (3, 15) into a forecast or a router, you would be overstating that line by 50%. The general lesson is narrow and practical: a price table is a versioned dependency with a verification date, not a constant you copy once. Announced changes can be cancelled before they take effect, so the rate that matters is the one in force today.
How to budget with two moving numbers
- Price your actual volume at the gateway rate. That is the number your invoice will reflect.
- Keep the list rate as the sanity check. If the two converge, the reseller stop is no longer doing anything for you.
- Check the note field before trusting a headline percentage. Promotional rates carry expiry dates; a percentage derived from one is only valid until it lapses.
- Re-check before you redesign a pipeline. Both sides move. A 24-hour batch pipeline built around a rate that has since changed is a sunk cost.
FAQ
Is the gateway always half of list price?
On 23 of the 24 models we can check, yes — exactly 50% below list on input. The 24th, gpt-5.6-sol, is 37.5% below on input and 25% below on output because OpenAI lists $4/$20 as promotional pricing valid at least through 2026-11-21. That single exception is the reason we do not state a flat rate anywhere on this site.
Which models can you actually verify?
24 of the 827 models in the catalogue carry a published list rate we can cite: 20 Anthropic models and 4 OpenAI. For the rest the gateway rate is the only figure available, and those pages say so rather than implying a discount that cannot be checked.
Do official prices change?
Yes, on the vendor's own schedule, and announced changes do not always take effect. Anthropic scheduled a rise on Claude Sonnet 5 from $2/$10 to $3/$15 for 1 September 2026 and cancelled it on 10 August, leaving the launch rate as the standard price. A price table with effective dates is a dependency with a version history, not a constant.
Why would the same model have two prices at all?
Because they are two different things. The list rate is what the model's own vendor charges you directly. The gateway rate is what a reseller charges for the same model, and it moves with upstream promotions rather than on the vendor's calendar. Neither is 'the' price — they are two quotes for the same work.
How should I budget when both numbers move?
Use the gateway rate for the volume you actually run, and re-check it whenever you change model or scale. Keep the list rate as the reference point for whether the gateway is still competitive. Do not hardcode either into a forecast without a verification date next to it.
Can I get these rates on every model?
Gateway rates, yes — the catalogue covers 827 models. List rates, no: most Chinese-lab models and most OpenAI variants carry no published USD list rate we can cite, so no comparison is possible and none is claimed.