Comparing AI API Prices: How to Read the Numbers Honestly
By Darus S Alfatah · 2025-07-23 · Review · 10 menit
Author note: Price per token misleads when a cheap model needs three tries. We compare per completed task.
Why Headline Prices Mislead
A price per million tokens looks comparable until you notice that input and output are billed at different rates, that long context may cost more, and that cached tokens get a discount. A model that is cheaper per token can be more expensive per completed task if it needs retries or produces longer answers.
The Four Numbers That Matter
- Input price — what you pay to read the prompt and context.
- Output price — usually two to four times the input price.
- Cache price — cheaper for repeated system prompts.
- Retry rate — the hidden multiplier nobody puts on a pricing page.
A Simple Comparison
Model A: input 200/million, output 400/million
Model B: input 250/million, output 250/million
For 2,000 input + 500 output:
A = (2000*200 + 500*400)/1e6 = 0.60
B = (2000*250 + 500*250)/1e6 = 0.625
In this example the cheaper input model wins, but reverse the ratio and the conclusion flips. Always compute with your own token mix.
Cost per Successful Task
The right unit is cost per task completed correctly, not cost per token. If a model fails ten percent of the time and must be retried, its effective cost is ten percent higher. If its answers are longer than needed, output cost rises. Measure both across a sample of real requests.
What Drives Your Bill
Context length dominates most applications. Teams that trim history, retrieve relevant sections, and cache repeated prompts typically cut costs by half without changing models. Model choice is the second lever; prompt discipline is the first.
Frequently Asked Questions
Is the cheapest model always best? No. Cheap models that need retries cost more than a mid-tier model that succeeds first time.
How do I estimate monthly cost? Multiply requests by average tokens, then apply prices; add a margin for experiments and retries.
Does streaming save money? Not directly, but it lets users stop early, which does save tokens.
Is batch pricing worth it? For non-urgent work, yes; it is usually much cheaper than real-time.
How often do prices change? Frequently. Review your model choices quarterly.
Conclusion
Compare with your own traffic, include retries and cache, and measure cost per successful task. That number is the only one worth negotiating over.
Reading the Invoice
Reconcile the provider invoice with your own logs monthly. Differences usually come from retries, failed requests that still consume input tokens, and experiments run by team members outside the tracked pipeline. Finding them early keeps surprises small and makes future estimates trustworthy.
Keep a one-page cost sheet per feature: model, average tokens, requests per month, and total. It is the simplest tool that prevents budget arguments later.
Negotiating and Switching
Once you know cost per successful task, switching becomes a calculation rather than a guess. Test the alternative on your own sample, compare quality and retries, then migrate only the features where the numbers clearly improve.
Finally, remember that price is not the only cost. Downtime, migration effort, and support quality also decide whether a cheaper provider is genuinely cheaper for your team.
Example: cost per task: Model A is cheap per token but needs two tries; model B is pricier but one-shot. In many cases, B is actually cheaper.
Frequently Asked Questions
What is the basis for this review?
The review is based on practical criteria such as output quality, speed, cost, and ease of integration for developers.
Is this review unbiased?
The review is written transparently based on experience and cited official sources; readers are encouraged to try things out for their own needs.
Can review results change over time?
Yes. Models and prices can change, so the review is updated when important changes happen.