You comparison-shop AI tools by sticker price. That's exactly what the vendors want.
Here's the thing nobody in AI sales will tell you: the price per token on the pricing page is almost irrelevant for small businesses. The real lever is something called cache economics — and understanding it can cut your AI bill by nearly half without switching tools.
When you send a prompt to an AI model, most of what you're sending is the same every single time:
A typical business prompt might be 10,000 tokens — but 9,000 of those are identical across every call. Only 1,000 tokens change each time (the actual question, the specific customer data, the new instruction).
The question that determines your bill: what does your AI vendor charge for those 9,000 repeated tokens?
Some AI providers recognize that repeated context is, well, repeated. They cache it — store a copy — so they don't have to reprocess the whole thing from scratch on every call. When they do, they charge you less for the cached portion.
Here's what that looks like with real numbers:
| Scenario | Full Price (no cache) | With Cache |
|---|---|---|
| Input: 10K tokens | $0.30 | $0.045 (cache read) |
| Output: 500 tokens | $0.15 | $0.15 (no change) |
| Per-call total | $0.45 | ~$0.20 |
Same sticker price. Same model. 55% cheaper per call because the vendor handles repeated context efficiently.
This isn't hypothetical. When Anthropic shipped Claude Fable 5.1, the sticker price stayed the same as Fable 5 — but the cache read price quietly dropped, cutting real-world costs roughly 45% for heavy users who send consistent context.
Enterprise AI teams have procurement specialists who model total cost of ownership. You probably don't. You see "$0.003 per 1K tokens" and make a decision.
But if you're running any of these, cache economics determines your real spend:
That's most SMB AI use cases. The repetitive ones. The ones where cache savings pile up fastest.
Before committing to any AI tool or API, ask the vendor these three questions:
If they say "we don't cache" or give you a blank stare, that's a red flag. You'll pay full price for the same 9,000 tokens on every single call. Good answers:
Red flags:
A vendor who understands your use case should be able to estimate this. If you're sending the same system prompt and business context every call, the answer should be 70-90%. If the vendor can't answer this, they're not optimizing for your cost — they're optimizing for their margin.
In September 2026, the AI code editor Cursor got banned from a major provider. Thousands of customers who relied on one company for both their AI memory and their AI files suddenly had neither.
The lesson extends beyond code editors: never let one provider hold both your memory and your files.
For small businesses, this means: Cache economics makes this practical: when you control your context (system prompts, business docs, templates), you can test whether a new provider handles it efficiently before migrating. You're not locked in.
Combine cache awareness with a two-tier model strategy:
Tier 1 — Routine work (cheap model, cache-heavy): Tier 2 — Judgment calls (capable model, cache still matters):
The beauty: your Tier 1 system prompt is probably 95% cacheable, and your Tier 2 context is probably 80% cacheable. Both tiers benefit from cache economics — but the savings on Tier 1 (where you make 10x more calls) is where the money adds up.
The AI market is consolidating and pricing is shifting fast. The businesses that understand cache economics will lock in savings while everyone else pays sticker price for the same service.
Want a template for auditing your AI vendor costs? The SMB Scale Up cost audit checklist walks through these questions with scoring — grab it from our resources.