/ Pricing Methodology

Pricing methodology

Sources · billing conventions · verification rules · corrections

This page explains where Reflex Hub's model pricing data comes from, under what conventions it is recorded, how often it is verified, and what happens when something is wrong. It complements the guide How to calculate LLM API costs: the guide teaches you to recompute numbers yourself, while this page documents the rules we use to compile and present the data.

1. Sources and collection principles

· Single source of truth: all model data lives in one models.json file; the model library, cost calculator, comparison pages and sitemap are all generated from that same file by the build script;
· Only official published list prices are recorded, sourced from official pricing pages or documentation (e.g. Alibaba Cloud Model Studio model pages, Tencent Cloud billing docs, OpenAI/Anthropic/Google official pricing pages). Every model carries a source name, a source link and a verification date, all clickable;
· Reseller quotes, second-hand reposts and community compilations are never used as a basis; after a vendor changes prices, the record is updated only after verification against the official page;
· This site is a third-party data compilation with no affiliation or endorsement from any vendor.

2. Billing units and original currencies

· Prices are recorded per 1M tokens, with input and output billed separately;
· Prices are presented in each vendor's original billing currency: CNY for Chinese vendors, USD for OpenAI / Anthropic / Google. A USD list price is never disguised as CNY;
· Cross-currency comparison and estimation use a static reference rate (the CFETS central parity rate), with the rate date shown and "indicative only" stated. When the reference rate is unverified, pages say conversion is unavailable instead of converting silently;
· Example monthly costs use fixed assumptions (100 calls/day, 2,000 input / 800 output tokens per call, off-peak rates, 30% cache hit rate, 30-day month), identical to the cost calculator defaults — recompute with your own usage anytime.

3. Input / output / cache prices

· Input and output prices are list prices on a cache-miss basis;
· Cache-hit prices are recorded separately; models without a published cache price are estimated at the standard input price (no cache discount is assumed);
· Cache write fees (e.g. OpenAI's 30-minute tier, Anthropic's 5-minute tier) are noted on the corresponding model pages but excluded from example monthly costs because they couple with retention duration and usage assumptions.

4. Context-length tiers and unpublished rates

· Some models bill by input-length tier (e.g. MiniMax M3 at ≤512K / >512K, Claude Haiku 5.5 at ≤100K / >100K). The calculator picks the tier from your per-call input length and labels it on the page;
· Tiers whose values the vendor has not published are marked "not published": estimates in that range return "cannot estimate" and never borrow the lower tier's price (e.g. GPT-6 series >272K long-context rates are unpublished — long-input estimation fails explicitly instead of returning a deceptively low number).

5. Time-based pricing

· DeepSeek models (via Tencent Cloud) bill peak/off-peak: the off-peak rate is about half the peak rate, with peak on weekdays 9:00–12:00 and 14:00–18:00;
· Example monthly costs default to the off-peak basis, matching the "Off-peak" toggle in the cost calculator; switch to peak to see both bases.

6. Verification-date semantics

· verifiedAt: the date both pricing and specs passed full verification; partially verified models keep an earlier verifiedAt rather than pretending to be "fully verified";
· priceVerifiedAt / specVerifiedAt: independent verification dates for prices and specs, which may be earlier than verifiedAt;
· contentUpdatedAt: the last substantive change to visible page content (prices, sources, specs, copy); it drives the sitemap lastmod. Re-verification without content changes does not update lastmod.

7. Unknown values and estimation limits

· Unpublished prices, rates and specs are always labeled "not published" and never filled with guesses;
· Estimates exclude free tiers, resource packs, subscription discounts and limited-time promotions (e.g. Qwen3.7-Plus shows a limited-time 20% discount on the official page; this site records the list price and notes the promotion without assuming an end date);
· Estimates are not provider invoices, do not constitute purchase advice, and are not used to infer model capability or quality rankings.

8. Reporting corrections

If a price, spec, source link or calculation disagrees with the official page, please open an issue:
github.com/cheneychen2000/ai-station-tools/issues
Include the page URL, the specific field and a verifiable official source link. Do not submit API keys or other sensitive data in public issues. Once verified, the data and its verification date are updated together.