Methodology
How AI Cost Estimator calculates its estimates, where the data comes from, and when each source was last checked.
Prices last verified between and against each provider’s published price list.See every source
LLM API estimates
Each request is split into six price buckets: standard input, cached input and output, and the batch version of each. With I input tokens and O output tokens per request, a share c of input read from the prompt cache and a share b of requests sent through a batch API, the expected tokens per request in each bucket are:
input = (1 − b) × I × (1 − c) batch input = b × I × (1 − c) cached input = (1 − b) × I × c batch cached input = b × I × c output = (1 − b) × O batch output = b × O
Each bucket is multiplied by the model's published price per million tokens and the results are added to give the cost per request. The daily cost is the cost per request times requests per day. A month is 365 ÷ 12 days and a year is 365 days.
- Price tiers. Some models charge more for long prompts. The tier is chosen from the average prompt length (cached and uncached input together), and that tier's prices apply to the whole request.
- Missing prices. A bucket with no tokens needs no price, so a model with no published batch price can still be priced when the batch share is zero. If your settings need a price the provider does not publish, or that we have not verified, the model is listed as not priced for that workload. We never fill the gap with an estimate.
- Comparison. The comparison table prices the same workload on every model (13 at present) and ranks them from cheapest to most expensive. Its list-price columns show the standard rates for the tier your prompt length falls in.
GPU rental estimates
For the chosen GPU, provider and pricing option, every matching offer is priced and the cheapest is used. Where a provider rents GPUs only in fixed instance sizes, the GPU count is rounded up to whole instances:
instances = ceiling(GPUs ÷ GPUs per instance) hourly cost = instances × price per instance-hour (or billed GPUs × price per GPU-hour) billed hours per month = hours per day × 365 ÷ 12 (reserved: 24 × 365 ÷ 12) monthly cost = hourly cost × billed hours per month
- Reservations. A 1-year reserved price is spread evenly over the 8,760 hours of the year and charged for every hour, used or not, because that is how a commitment is paid.
- Spot prices change often and the provider can reclaim the GPUs at short notice. Stored spot prices are a dated snapshot.
- Cost per million tokens. Tokens per month = GPUs × tokens per second per GPU × busy share × 3,600 × hours switched on per month. The cost per million tokens is the monthly cost divided by that figure. Throughput comes from a published benchmark where we have a verified one, or from the figure you enter. Benchmarks measure maximum throughput for one model, so the result is a best case. Without a throughput figure, no cost per token is shown.
Of the 6 GPU types in our data, 1 currently has a verified benchmark figure.
Currencies
Every provider in our data publishes prices in US dollars, and all calculations are done in US dollars. US pages show the results as they are. UK pages divide them by the Bank of England daily spot rate for US dollars into sterling (series XUDLUSS), currently £1 = US$1.3403 for . The rate is shown with every UK result. If no verified rate is available, UK pages show no pound figures rather than an assumed rate.
What the estimates leave out
- Fees for writing to a prompt cache and for cache storage. Cache writes are charged at the standard input price.
- Tool, web search, image and audio charges, and embedding calls.
- Differences between tokenizers: the same token counts are applied to every model.
- Priority, flex and other processing options beyond standard and batch, and free tiers.
- Storage, networking and data transfer for rented GPUs, and the engineering time to run them.
- Discounts negotiated directly with a provider.
How prices are checked
Every figure is read from the provider's own pricing page or documentation and stored with the exact page it came from and the date it was checked. Prices are used exactly as the provider publishes them. When a provider announces a dated price change or a promotion, we note the date and recheck the figure once it takes effect. The table below lists every source with the date it was last checked.
Data sources
40 data points are awaiting verification. Until verified, they are not used in any result.
Limitations
Estimates depend on your inputs and on averages from published sources. Real prices vary by provider, location and timing. Treat results as a starting point, not a quote.
Corrections
If you find a figure that is wrong or out of date, please let us know.