How we observe MaaS platforms

1. Observations and official announcements

We periodically fetch official model lists, pricing pages, release notes and blogs, then compare them with the previous snapshot. These are observations of page text: truncation, page movement and recomputation can introduce noise.A page change is not an official release, retirement or price announcement; that requires an official announcement.

2. Five distinct timestamps

• First observed (observedAt): when our collector saw the content;
• Source publication time (publishedAt): the supplier’s stated publication time; empty without a source, never substituted with the observation time;
• Price observation interval: the observation interval and version history of the price;
• Data through (dataThrough): the dataset coverage boundary, currently 2026-11-01；
• Site build time: when the site was published, distinct from source data time.

The home and pricing pages show their own data timestamps, rather than the site build time.

3. Evidence completeness

Each pricing fact includes a saved excerpt, locator and excerpt hash. Evidence has three completeness states: complete (the model, amount and unit can be checked in the excerpt), partial (the evidence structure exists but an element cannot be checked in the excerpt; existence alone is not verification), unavailable (historical evidence cannot be recovered). Summaries, locators and official sources support different levels of verification.

4. Generated summaries can be wrong

Some observation summaries are generated by an LLM (summaryOrigin=llm) and may be wrong. Verify through: Stable ID → structured facts → saved excerpt → official source. Generated RSS and API summaries disclose their origin.

5. Complete billing conditions

Each price retains its currency (USD/CNY), unit, input/output/cache component, region, tier, billing mode and additional conditions such as context bands and peak/off-peak periods.We do not infer a lowest price across incompatible units or conditions. Candidates remain separate so you can assess their applicable conditions.

6. Supplemental datasets

Usage rankings come from third parties; exchange rates and some capability rankings are manually maintained. Each shows its own update time. Unknown values remain unknown.

7. Three distinct health signals

• Fetch succeeded: the page was downloaded and parsed;
• Pricing validation passed: amounts, evidence chains and identities passed validation;
• Last accepted fact: the most recent accepted fact from this source.

These signals are distinct. When an update fails and previous data is retained, the state is stale (with the last successful time), rather than fresh data. partial, unknown and unavailable retain their literal meanings.

8. Revisions, withdrawals and caches

Records support revisions and withdrawals while retaining history and reachable detail pages. RSS is a static projection; revisions preserve identity, but cached withdrawn entries cannot reliably be recalled. Detail pages and API responses are authoritative for current record status. For interface changes, see the changelog。

Language preference and country lookup

Only the homepage entry selects a default language by country: Chinese for mainland China (CN), English elsewhere. Your manual choice takes priority and is stored locally for 180 days. Deep links retain their language. Failed lookups fall back to your browser preference; without JavaScript the homepage stays Chinese. Use the language links to correct VPN or database inaccuracies.

Country lookup uses a local server database without adding a third-party address lookup or raw IP tracking. A database more than 21 days old is disabled and the fallback is used. This product includes GeoLite2 data created by MaxMind; distributed by geoip-country . The database is governed by GeoLite EULA and CC BY-SA 4.0 .

Feedback

Contact me: zhangchen3508@gmail.com。