How it works
How LLM Pricing Works
LLM Pricing normalizes public pricing data into a consistent format so you can compare models across providers. Here's how the data flows from source to estimate.
Data built Aug 10, 2026. Verify with providers before decisions.
Data collection
Primary source
The model catalog starts from LiteLLM's open-source metadata, which aggregates provider-published pricing. This gives us a consistent base across 139+ providers and 2,900+ models.
Manual overrides
When provider-specific pricing semantics need clearer labels, notes, or normalization, fields are manually reviewed or overridden. This catches edge cases that automated extraction misses.
Price normalization
Raw pricing data comes in different formats across providers. The normalization pipeline converts everything into consistent per-1M-token fields:
Input price
Cost per 1 million input tokens. This is the text you send to the model (prompts, context, documents).
Output price
Cost per 1 million output tokens. This is the text the model generates back. Output typically costs 3-4x more than input.
Cache pricing
Cached input price (usually 90% discount) and cache write price. Derived from base input pricing using provider-specific rules.
Cost estimation
The calculator uses a straightforward formula:
Monthly cost = (input_tokens × input_price + output_tokens × output_price) × monthly_requests
All prices are per 1 token (converted from per-1M-token). Cache hits use the cached input price instead of the base input price.
Benchmark data
Benchmark scores come from official model cards, launch posts, papers, and benchmark operators. Each score is linked to its source. Coverage is intentionally explicit — if a model doesn't have a verified benchmark score, it shows as blank rather than guessed.
What's not included
Not financial advice
All pricing figures are for reference only. Actual invoices may differ because of caching, discounts, regions, batch processing, usage tiers, and provider-specific billing rules.
No backend or database
The site is fully static. All calculations happen in your browser. No data is sent to any server. The model database is generated periodically and served as JSON files.
Want to verify the data?
Check the data sources page for provenance details, or download the raw models.json to inspect the data yourself.