If your OpenAI invoice jumped and you're trying to figure out why before it happens again, the good news is it's almost always one of a handful of explainable, fixable causes — not a mystery. Here's what actually drives an unexpected OpenAI bill increase, in the order we see it most often.
Reasoning tokens are billed as output tokens — even when you never see them
If you're calling a reasoning model, the model's internal reasoning tokens aren't shown in the response, but OpenAI's own documentation is explicit that they "still occupy space in the model's context window and are billed as output tokens." Output tokens are the expensive half of any API call, so a jump in reasoning effort — or a model reaching for more thorough reasoning on a harder prompt — can inflate your bill even when the visible answer looks about the same length as before. Lowering the reasoning.effort parameter or setting a max_output_tokens cap are the two most direct levers.
Your account moved into a higher usage tier
OpenAI automatically raises your account through usage tiers as cumulative spend crosses set thresholds, unlocking higher requests-per-minute and tokens-per-minute limits at each step. That's not a price change by itself, but it does raise the ceiling on how much your app can actually do — so if a queue that used to throttle itself against a low rate limit suddenly has headroom to run at full speed, the resulting spend increase can look sudden even though nothing about your code changed.
Retries after a 429 error can quietly double- or triple-charge you
Hitting a rate limit returns an HTTP 429 with a Retry-After header. If your client — or a library you depend on — retries aggressively without honoring that header or backing off exponentially, you can end up sending, and paying for, the same logical request more than once. OpenAI's own SDKs handle this correctly by default, but a custom retry loop or an older integration might not.
Prompt caching only pays off if you're actually getting cache hits
Cached input tokens carry a steep discount versus standard input pricing, but that saving only applies on a cache hit. If your system prompt, tool definitions, or leading context change on every single call, you never qualify for the cached rate — you're paying the full uncached price while assuming, because caching is technically "enabled," that you're on the cheap path.
OpenAI changed something and didn't make a lot of noise about it
OpenAI does cut prices, raise them on specific models, rename processing tiers, and deprecate older models onto newer ones with different per-token rates — and none of that always comes with a prominent announcement. This is the exact gap VendorSpies is built to close: check the OpenAI calculator above for the latest verified rate and price history, and add a free watch so the next change reaches you before your invoice does.