True LLMs / TRUST CENTER
How we calculate this
Every number on True LLMs is either a sourced fact, a formula applied to sourced facts, or a clearly labeled estimate. Here's exactly how.
01 / Three kinds of number
Fact: Taken directly from a product's own public pricing/docs page, unmodified. Shown with a source link.
Derived: Calculated from facts using a fixed formula (e.g. credits ÷ price). Always shown with "See the math" expanded to the exact arithmetic.
Estimate: Depends on usage assumptions we don't control (prompt length, model mix, caching behavior). Always prefixed with "≈" and never presented with the same weight as a fact.
02 / Subscription value
How many dollars of credit a plan grants per subscription dollar.
Only computed for credit-metered plans with a disclosed monthly credit figure. Flat subscriptions, pay-as-you-go and bring-your-own-key plans don't have a comparable ratio and are shown as "Not available," not estimated.
03 / Request cost
What one request costs, from a model's per-token rates and a token-usage profile.
04 / Estimated requests
Where a product discloses a hard request quota instead (e.g. GitHub Copilot Free's 2,000 completions/month), we show that quota directly as a fact, not this formula.
05 / Deals
A model has at most one active deal applied at a time. Deals never stack.
- Discount: rate × (1 − discount%)
- Multiplier: rate ÷ usage multiplier
- Free: rate = $0
A deal is only shown as active if it's marked active AND (permanent, or the current date falls within its start/end range). Expired deals disappear automatically. Nothing expires by someone remembering to update a flag.
06 / Recommendations
The recommender is rules-based, not AI, and product-agnostic. It never checks which product or plan it's looking at by name.
- Sort every candidate plan by subscription price, ascending.
- Drop plans that don't meet stated requirements (premium models, long context).
- Walk the remaining plans in price order; recommend the first whose disclosed capacity covers your estimated usage.
- If none fit, recommend the highest-capacity plan available and say so explicitly.
This means the recommender can and does suggest a cheaper plan, or a different product entirely, whenever the numbers point that way.
Usage estimates vary with model choice, prompt size, output size, context growth, caching and task complexity. These are estimates, not guarantees.