How we source, check and calculate

Sources

We only take prices from a provider's own official pricing page or documentation. We don't use aggregators, resellers or third-party blog posts. Every row in every dataset records the source URL and the date we fetched it (capturedAt). If a page won't load or a price is ambiguous, we leave the row out rather than estimate it.

Confidence labels

• verified: the listed price appears unambiguously on the source page for the stated conditions.
• conditional: the price depends on tiering, a marketplace or a calculation we made, such as dividing an 8-GPU instance price by 8. The conditions column says what.
• disputed: we found conflicting information and are checking it.

Normalization

LLM API prices are in USD per one million tokens at the standard tier. Batch, volume and committed-use discounts are excluded. GPU prices are on-demand prices per GPU-hour. When a provider sells only multi-GPU instances, we divide the instance price by the GPU count and say so in the row. We show currencies as listed and don't convert them.

Freshness

An automated check flags any observation older than 30 days. A stale price is re-checked or marked as stale. It is never silently kept as current.

Break-even calculator formulas

• API cost = (fresh input × input rate + cached input × cached rate + output × output rate) ÷ 1,000,000
• Useful GPU-hours = (input tokens ÷ prefill tok/s + output tokens ÷ decode tok/s) ÷ 3,600
• GPUs = max(minimum GPUs, ⌈useful GPU-hours ÷ utilization ÷ 730⌉)
• Self-host cost = GPUs × 730 h × price per GPU-hour + fixed overhead
• Break-even: the smallest multiple of your workload, at the same input/output mix, where self-host cost ≤ API cost.

A month is 730 hours (8,760 ÷ 12). We assume GPUs are rented always-on. The model ignores taxes, egress, storage, autoscaling and spot capacity. It also doesn't check whether a GPU has enough memory for your model; that's what the minimum-GPU input is for.

What we don't do

We don't publish throughput figures we haven't measured, and we don't describe a provider as secure, compliant or certified without a primary source that states the scope.