Built for agents that are not allowed to guess

No. It means the phrase and the number both appear on that page, close together. Two measured calls against the same Anthropic pricing page prove the limit: 'Claude Opus 4.1 input is $15 per million tokens' returned supported at confidence 0.9, and 'Claude Opus 4.1 output is $30 per million tokens' also returned supported at 0.9, because $30 sits in the same price row even though it is not the output rate. The payload says it plainly in its own note: this is a source-evidence check, not a truth oracle. Always read the evidence snippets, which show exactly which text was matched.

It depends on the parse surface, and the response tells you which one was used. On providers parsed with html_proximity, normalization_notes states that input is the first price found and output is inferred from a 1.5x-to-8x band, which can pick the wrong column on a page that lists cache and batch tiers side by side. Every row also carries prices_found_usd with all the nearby rates, so on a measured Opus 4.1 row you can see the full set 15, 18.75, 30, 1.5 and 75. Bill from prices_found_usd and source_url, not from the single output number.

Because a provider's pricing table lists one row per tier, and the parser keeps them all rather than picking a winner. A measured comparison filtered to 'flash' returned gemini-3.1-flash repeatedly, at four different input rates: 0.125, 0.25, 0.45 and 1.5 per million tokens, which are the context and modality tiers listed on that page. cheapest_input and cheapest_output pick the lowest row, so they will report the cheapest tier of a model and not its standard rate. De-duplicate on model plus price before you show anything to a user.

You get an honest empty answer rather than a guess. A measured official_source call for 'some-random-startup-xyz' returned ok:true with provider null, matched false, official_url null, an empty all_sources object and a suggestion listing the eight supported keys. The disclaimer in the payload states that finding the official source for any entity on the internet is deliberately not claimed, and that the whitelist only grows with maintainer sign-off.

The page is fine. verify:true on official_source runs a lightweight reachability probe that can report false on hosts that answer redirects unusually. A measured source_url call on the very same URL, https://openai.com/api/pricing/, returned accessible true with status_code 200 after two hops. When reachability matters, use source_url, which gives you the full redirect chain with a status code per hop.

It follows the redirect chain to the end, including any locale rewrite the site performs on your request. The measured chain from https://openai.com/api/pricing/ was 308 to https://openai.com/business/pricing/#api, then 302 to https://openai.com/tr-TR/business/pricing/, which returned 200. Read final_url and final_domain rather than the URL you sent, and note that official.is_official is decided on the final domain, which is why that redirected page still scored true for openai.

It means the page published neither a date nor a usable archive snapshot. In a measured call, metadata.published_at, updated_at, title and canonical were all null, archive.available was false, and staleness fell back to 'unknown'. Modern marketing pages frequently ship no dateable metadata at all. When you need a timestamp on a price, use last_seen_at from pricing_lookup, which records when the page was actually read, rather than trying to infer the page's own age.

Use official_source, which resolves them normally, and skip pricing_lookup for those three. A measured pricing_lookup on aws returned ok:true with pricing_available false, an empty models array, and normalization_notes explaining that model pricing there lives inside compute-SKU pages with no stable structured slice, so it is out of scope. The official pricing URL still comes back, along with why_official showing the domain match, its age in days and the registrar.

Source Verification API is a ReefAPI endpoint group for source verification It returns live JSON through POST requests under /agent-source/v1.

Yes. ReefAPI starts with 1,000 free credits, no card required. Source Verification calls use the same shared credit balance as every other ReefAPI engine.

No login to Source Verification is needed for the API response. You call ReefAPI with your x-api-key header, and the playground can run live examples before you create a production key.

The page example is captured from a live source_url call, and production requests fetch live data through ReefAPI rather than a static sample.

Source Verification actions currently cost 1-3 credits per successful call. Failed or blocked calls are free, and all APIs draw from one credit pool.

Yes. Connect ReefAPI once through MCP and your assistant can call agent-source actions with the same key, credit pool and JSON envelope used by normal REST requests.