In plain English
- What happened
- OpenAI and Anthropic both publish per-token pricing that is low and getting lower with each model release, yet many small businesses see their monthly AI invoice go up anyway. The reason is that a bill is price per token multiplied by tokens used, and the volume side of that equation often grows faster than the price side shrinks.
- Why it matters to your business
- A business owner who only checks the headline price cut is comparing the wrong number to their invoice. The actual driver is usually how many tokens a tool burns per task — through longer conversations, repeated context, hidden reasoning, or multi-step agent calls — and that number is invisible unless someone goes looking for it.
- Do this
- Open the usage dashboard for whichever provider sends the bill and sort by token count rather than dollar total, to see whether volume or rate is actually moving.
The per-token prices OpenAI and Anthropic publish for their models are genuinely low, and both list current rates on their own pricing pages, alongside Anthropic's published rates. But an invoice is not the price per token. It is the price per token multiplied by how many tokens got used that month, and that second number is the one that quietly climbs while nobody is watching it.
So a bill can rise in the same month a headline price cut lands, because whatever tool sits on top of the model is doing more work per request than it used to: longer conversations, more attachments, more back-and-forth steps to finish a single task.
Where the extra tokens come from
One source is reasoning itself. According to OpenAI's reasoning models guide, newer models generate internal reasoning tokens before producing a visible answer, and those tokens are billed as output tokens even though they never appear in the reply a customer or staff member sees.
OpenAI's own documentation confirms that reasoning tokens are billed as output tokens, even though they are not included in the response text.
Another source is repeated context. Both companies' prompt-caching documentation exists specifically because, without it, the same system instructions or conversation history get sent and billed as fresh input tokens on every single call — described in OpenAI's prompt caching guide and Anthropic's prompt caching documentation. An automation that calls a model five times to finish one task can end up paying for that history five times over, not once.
Caching can bring part of it back
Prompt caching is designed to discount those repeated tokens, and both providers document it as a real cost reduction. The catch is that it is not always automatic: Anthropic's documentation describes it as something a developer has to set up with specific markers in the request, so the saving only applies if whoever built the integration actually turned it on.
Reality check
Switching to a cheaper-sounding model name rarely fixes a bill that is rising because of volume rather than rate. If the same workflow is calling the model more often or feeding it more context than before, the cost comes back under the new model too — the fix is in how the tool is used, not which model it points to.
- 1
Pull the usage breakdown, not just the total
Open the usage dashboard for the provider being billed (OpenAI's usage page or Anthropic's console) and look at token counts by day and by model, rather than the single dollar figure on the invoice.
- 2
Check for reasoning tokens
If the workflow uses a reasoning-capable model, look for a separate reasoning-token line in the usage breakdown — this is often the largest hidden share of the bill.
- 3
Ask whether caching is turned on
Check with whoever built or configured the integration whether prompt caching is enabled; if it was never set up, repeated context is being billed at full price every call.
- 4
Count the calls per task
Find out whether an agent or automation makes several model calls to complete one customer request, since each call can resend the same context and multiply the token count.
- 5
Compare token volume month over month
Track total tokens used over time, not just the dollar total, since a falling per-token rate can still sit under a rising bill if volume is growing faster.
Sources
