How we validate AI API pricing data

Data pipeline

1. Provider pricing is reviewed for an exact API model, billing channel and service tier. A dated evidence record stores the rates and conditions.
2. OpenRouter is a discovery source for routed offers. It cannot overwrite direct-provider prices, invent cache discounts or verify model identity through fuzzy names.
3. The refresh reads each official pricing page it knows how to parse. When that reader misses an offer, a LiteLLM row may fill it if the provider, model id, and tier are the same. A matched rate is written into the catalogue and that offer’s check date is renewed for seven days. An offer that still cannot be read keeps its previous rate and date.
4. Build checks reject missing evidence, price drift and unsupported trust claims. Evidence has a seven-day review window; stale and unreviewed offers are excluded from default recommendations.
5. Download the dated catalogue to inspect per-offer review status and scope.

What each dataset is for

Workload estimates (input, cached input, and output $/MTok, including Azure, Bedrock, and Vertex host prices) come from the MCC validated dataset. Intelligence Index, output speed, and cost-per-task on the explorer come from Artificial Analysis and are attributed there. When a calculator model is mapped to an AA slug, the explorer shows the same token prices as the calculator. Benchmark fetch dates and price-review dates are independent. The score belongs to the named benchmark variant, including its reasoning effort; it is not an intrinsic guarantee for every API configuration. Missing scores remain unavailable.

Cost formula

Monthly estimate = (non-cached input tokens × input $/MTok) + (cached input tokens × cached $/MTok when published) + (output tokens × output $/MTok), scaled to messages/month × tokens/message. If a model has no cached rate, cached tokens bill at the normal input rate.

Confidence tiers

• Source reviewed — exact offer checked against a dated primary source, within the stated billing scope.
• Primary source review does not mean a second source confirmed the price or that every billing dimension is included.
• Needs review — missing, stale or conflicting evidence; excluded from recommendations.

Why some offers are excluded

The September 23 review covered all 59 previously excluded offers: 26 restored, 27 archived as retired, deprecated or duplicate entries, 5 requiring other billing inputs, and 1 with unresolved primary pricing evidence. Archived rates are historical records, not current purchase options.

Not included in estimates

The UI estimates token charges. Cache writes and storage, tool-call fees, web search grounding, vector storage, image tokenization, regional multipliers, taxes, and enterprise discounts are not modeled. Add those from your provider console when they apply.

Canonical tool vs article

The interactive calculator lives at modelcostcomparison.com. A long-form article with additional context remains on lazige.agency and points here as the canonical tool URL.