A 33x Price Gap for the Same Model Is Not a Market. It Is a Tax on Not Checking.

Key takeaways

• Daily price tracking surfaced a 33x spread on the same model across providers, as reported in the Hacker News discussion.
• A gap that large is not explained by quality. The weights are the same. It is explained by attention.
• Nobody is paying 33x on purpose. They are paying it because the endpoint was chosen once, early, and never revisited.
• Switching cost is mostly psychological now that most endpoints are API-compatible.
• Price is a moving number. Treating it as static is the actual mistake.

What was found

The author tracks LLM API list prices daily and reports finding a 33x cost gap for the same model across different providers and endpoints. Not a comparable model. Not a smaller variant. The same model.

Most commentary about AI costs is about which model to use. This finding says something less comfortable: for a meaningful slice of teams, the model choice was fine and the bill is still several multiples too high.

How is a 33x gap even possible?

Because list price for a hosted model is not a cost calculation, it is a positioning decision. Providers running the same open weights differ on hardware, batching, quantization, region, committed capacity, and how much margin they want. A provider selling inference as a loss leader to win platform adoption and a provider selling it as a premium managed service will land far apart even with an identical model behind the endpoint.

Add two more multipliers on top of the list rate. Prompt caching discounts differ by provider. Batch or asynchronous tiers, where offered, are usually cheaper than realtime. Stack those and the effective spread widens further than the sticker spread alone.

Why do teams keep paying the high end?

Here is the uncomfortable part, and the observation the source does not make: the 33x gap survives because AI spend is usually nobody's job.

Infrastructure spend has an owner. Cloud has a FinOps function, a tagging policy, a monthly review. LLM spend is new enough that it typically sits with whoever shipped the feature, gets approved as a single line item, and never gets re-tendered. So the provider chosen during the prototype, when spend was $40 a month and switching felt pointless, is still the provider at $40,000 a month, when switching is the highest-leverage hour of work available to the company.

That is not a technical failure. It is an ownership gap.

What would a 33x gap actually cost you?

Use an illustrative example rather than a promise, since your token mix decides the real number. Say a product burns 50M input and 10M output tokens a month. At the cheap end of a wide spread that might be a few hundred dollars. At the expensive end, the same traffic is a five figure line item. Same model, same output quality, same users.

At seed stage that difference is your entire infrastructure budget. At Series A it is a hire. The point is not that everyone is overpaying 33x, it is that you cannot know where you sit without checking, and almost nobody checks.

What should you actually do?

Four steps, none of them heroic:

1. 1Pull your last invoice and convert it to cost per active user, not total spend. Total spend hides everything.
2. 2List every endpoint serving the models you use, and price your real token split against each one.
3. 3Model caching and batch tiers separately. They are where the effective gap widens beyond the list gap.
4. 4Put a recurring calendar entry on it. Quarterly is enough. The prices move.

If that sounds like procurement rather than engineering, that is because it is, and that is exactly why it does not get done.

A 33x spread is not a market signal, it is an unclaimed refund. You can price your own token split against multiple providers in the Calcaas LLM cost calculator and find out which end of the gap you are on.