AI cost: price per token isn't enough to choose the model
In a customer service, research or analysis application, the company doesn't buy tokens. It wants a correct answer, a usable document or a task solved on time. Between the list price and that result come the accuracy rate, the attempts, the resent context and human review. That's the difference an AWS benchmark, published on September 11, put at the center: cost per correct answer, per agent trajectory and per approved deliverable.
The math changes when the result goes into the formula
In the AIME test AWS describes, the observed cost per correct answer was US$ 0.0021 for gpt-5.6-luna and US$ 0.0139 for gpt-5.4-mini in the configuration evaluated. In another set, the model with the lower nominal price didn't have the lowest cost per approved answer because its approval rate was also lower.
These numbers aren't a universal price list: they come from a sample with specific models, regions, configurations and criteria. The value is in the question the method makes explicit: how much does it cost to get a result the company accepts?
Agents make the trajectory part of the bill
In flows with tools, the model can make several calls and resend instructions, results and documents. AWS observed, in a research task with 50 questions, that the number of turns changed the volume of tokens and the cost of an approved answer. So the calculation needs to include latency, failures, review, rework and integration, not just the price of one call.
The company's benchmark should start from a real process
Before switching models, leadership can organize a small test:
- select 50 to 100 real tasks, with anonymized data and an expected result;
- define what an approved answer is and what requires review or discarding;
- measure tokens, calls, time, failures, corrections and supervision;
- compare the models on the same task and repeat the test when price, volume or flow design changes.
The goal isn't to choose the most technically advanced model. It's to find out which configuration meets the required quality level without shifting the cost onto the team that reviews, corrects and maintains the process.
The economic decision depends on the quality bar
The AWS benchmark has clear limits: small samples, different configurations across platforms and automatic evaluators in part of the tasks. Close differences should be treated as directional and validated in the operation.
This matters because acceptable cost depends on risk. A preliminary email classification can tolerate a higher review rate. A financial summary, a regulatory document or a recommendation that affects a customer needs a higher quality bar.
Price per token is still useful information. It just shouldn't take the place of the decision on its own. The more mature question is which process will be best served, with what quality and at what total cost.
To start, pick a repetitive routine, define a baseline and track the cost per approved result for a few weeks. The savings show up when the measure follows the real work.
Measure the cost of the complete task
Define what counts as a correct answer, an approved deliverable and a completed task. Compare models with the same set of cases, record failures and review time and include usage, integration and supervision cost in the decision.
Sources
AWS, 11 Sep 2026. https://aws.amazon.com/blogs/machine-learning/beyond-the-price-per-token-choosing-the-right-openai-model-on-amazon-bedrock-for-your-workload/