Yeah, the haggling is real. I'm trying to budget for API costs for a new project, and it's impossible when the real number is "call us." It's like there's a whole hidden economy.
Does anyone know if the smaller, newer AI providers work the same way, or do they actually charge the list price?
That "call us" quote is so frustrating when you're trying to build a business case.
In my experience, the smaller AI startups often do stick closer to list price, at least for their public APIs. But the moment you need any custom volume or support, it shifts to the same hidden economy. I've seen list prices jump 2-3x when you move from their public tier to an "enterprise" plan, even if the usage is the same.
Is there any pattern to which tools let you build a real budget from their docs? I feel like it's getting harder.
Still learning.
You've hit on the core of the problem: the pivot from a transparent, usage-based model to an opaque "enterprise" one. The pattern I've observed is that tools with a purely consumption-based pricing model, like many cloud data services (e.g., Snowflake, Databricks, certain AWS services), tend to allow for more accurate budgetary modeling from their public docs. Their list price is often the real price, minus eventual volume discounts.
The moment a vendor introduces "support," "SLA," or "professional services" as a bundled line item, predictability vanishes. I recently benchmarked two observancy platforms. One had a public price per gigabyte ingested. The other required an "enterprise" quote. For comparable volumes, the quoted enterprise price was 4.2x higher than extrapolating from the competitor's public rate card, with the difference attributed to "dedicated support" and "custom integrations" we didn't request.
So the heuristic is: if the pricing page has a calculator and bills solely on measurable units (GB, hours, API calls), you can build a model. If the pricing page ends with "Contact Sales," your model is fiction until you negotiate. Sadly, the latter is becoming the default for anything labeled "platform."
—chris