Home › Guides › How we verify prices

How LLMPrices verifies prices — and where price data goes wrong

Step 1: one source, pulled whole, every day

Once a day we pull the complete public catalog from OpenRouter's API and store it as a dated snapshot. Every page on the site — all 462 model pages, the rankings, the comparisons — is regenerated from that single snapshot, and the snapshot date is printed on each page. Nothing is edited by hand: if a figure looks wrong, the fix is tomorrow's snapshot, not a manual patch. Snapshots accumulate, which is what lets the price changes page show exact day-over-day moves instead of vibes.

Step 2: normalization, with no silent estimates

Providers quote per-token rates; we multiply out to US dollars per 1M tokens for input, output and cached input so every model is comparable on one scale. Where a provider does not publish a rate (some cached-input prices, some context limits), the field is shown as "—". An honest blank beats a plausible guess, because a guessed price is worse than no price: it gets copied into budgets.

Step 3: the official cross-check

Catalogs are convenient and usually right — but "usually" is doing work in that sentence. For 77 flagship models from eight major providers (OpenAI, Anthropic, Google, DeepSeek, xAI, Mistral, MiniMax and Moonshot AI), we read the prices from each provider's own pricing page (last full check: 2026-10-21) and publish the comparison on the official prices page and on each model page. Agreements build trust in the catalog; disagreements get a flag, not a cover-up.

Where price data goes wrong: five failure modes

1. Catalog lag. A provider announces a cut on its blog; aggregators and trackers update hours to days later. Any price without a visible "as of" date should be treated as history, not fact.

2. Source switches. Sometimes the number changes because the feed changed, not the price. In October 2026 we watched one flagship's catalog price jump from $0.079 to $0.70 per 1M input tokens and back within a single week — a data-source switch, not a 9× price rise and fall. Wild day-over-day swings on the price changes page should be read with this in mind: real provider cuts are usually announced; feed artifacts are not.

3. Catalog ≠ official. The cross-check regularly finds models listed in the catalog below the provider's official list price, and occasionally a model with a free catalog route while the official page charges for it. Both can be real (reseller margins, promotional routes), but you should know which number you're budgeting against.

4. Renames and variants. A model that "disappears" from a catalog is often just renamed, and variants (batch editions, regional editions, dated snapshots of the same model) fragment listings. Trackers that don't sweep dead pages end up comparing ghosts.

5. Time-shaped prices. Some providers run peak/off-peak dual pricing or promotional windows with an expiry date. A single list price scraped at the wrong moment misrepresents both. Our notes track known promo windows for flagships, but the general lesson stands: for a big commitment, confirm the rate — and its expiry — on the provider's page.

What we do when sources disagree

Three rules. Publish the catalog number (it's the consistent, daily series). Show the official number next to it wherever we have one. Flag the difference visibly. We never silently average two sources into a third, wronger number, and we never overwrite today's figure with yesterday's because it "looks off" — the audit trail is the price changes page plus the snapshot date on every page.

How to verify us

Don't take this guide's word for it. The official prices page links each flagship to its provider source; the JSON API exposes the entire table with its snapshot date, so you can diff our numbers against any other source programmatically; and the full method, including limitations (list prices only — enterprise discounts and regional pricing are invisible to any public tracker), is on the methodology page. If you find a figure that's wrong, the corrections address is on the about page — and because of Step 1, a confirmed error is gone within a day.

Start over: how LLM API pricing actually works →

By the LLMPrices editorial team · October 2026. All pricing guides →