Correction — October 10, 2026. The historical comparison below comes from the 2025 AI Index, covering November 2022 to October 2024. Our previous version labelled it as the 2026 report and understated the elapsed period. Unsupported hardware and training-cost comparisons have been removed.
A historical comparison, with its conditions
Stanford’s 2025 AI Index reports a decline from roughly $20 to $0.07 per million tokens for models achieving GPT-3.5-level performance on MMLU (64.8%), between November 2022 and October 2024. That is more than a 280-fold reduction for the benchmark-matched comparison. It is not a quote for every model, a current provider tariff or a promise that your application’s total cost fell by the same factor.
A benchmark can help control one aspect of capability. It does not control your prompt length, retry rate, required response time or acceptance criteria. Save the original price date and benchmark threshold when citing this comparison.
Calculate your own accepted-result cost
This calculator runs locally in your browser. Replace the hypothetical defaults with rates and usage from the service you actually use. Use one currency for both rates. It estimates token charges only; it excludes human review, hosting, tools and other service fees.
Reproduce the calculation
Total token cost = attempts × (input tokens × input rate + output tokens × output rate) ÷ 1,000,000.
Expected accepted outputs = attempts × acceptance percentage ÷ 100.
Cost per accepted output = total token cost ÷ expected accepted outputs.
With 100 attempts, 2,500 input tokens, 800 output tokens, hypothetical rates of 1 and 4 currency units per million tokens, and 80% acceptance: input costs 0.25, output costs 0.32, and the total is 0.57. Dividing by 80 gives 0.007125 per accepted output. Expected outputs may be fractional when you use an estimated percentage; actual accepted counts are integers. These are assumptions, not measured service results.
If no output is accepted, the accepted-result cost is undefined. The calculator asks for a positive acceptance percentage rather than presenting a false zero. In a real run, log the actual accepted count and record failed attempts before comparing alternatives.
Use the result to make a product decision
Choose an acceptance rule before testing: for example, a generated record must have all required fields, cite the supplied evidence and pass a human check on a small audited sample. A fluent answer is not automatically accepted. Keep the same task set when comparing two models.
| Record | Decision it supports |
|---|---|
| Model version and rate date | Prices and behaviour can change; retain the version actually tested. |
| All attempts and billed token classes | Retries and long outputs can change the apparent saving. If cached or reasoning tokens have separate rates, calculate them separately. |
| Accepted count and failure reasons | A cheap call can be expensive per useful result. |
| Review time and infrastructure cost | Add them to total spending when you need a full production estimate. |
| Latency and operating constraints | Evaluate these separately; token price is only one part of the decision. |
A sensible next step is a small, reproducible trial with your own task set. Treat an estimated saving as a hypothesis until the trial records useful outputs, actual charges and the work needed to correct failures.
Sources and scope
- Stanford HAI, The 2025 AI Index: State of AI in 10 Charts: Historical benchmark-matched inference-price comparison; not a current API price list.
- 2025 AI Index, Chapter 1: Detailed context for the model-performance and inference-cost comparison.
AI-assisted drafting. The linked material supports the stated facts; the examples and worksheets are Luminesca’s own editorial work. We did not conduct facility measurements, commercial project due diligence or model benchmarking for this article. Send a correction with the specific passage and supporting source.