How the LLM API price tracker works
Updated
BroadRouter's LLM API price tracker holds 1,671 list prices for 430 models from 41 providers. Each price is read from the provider's own pricing page with AI assistance and stored with its source URL, date checked and a verbatim quote. A script re-checks the source pages weekly; the newest check is 2026-11-15. BroadNet, which runs this site, is listed separately and never ranked.
Disclosure: BroadRouter is BroadNet's own site and links to BroadNet Inference, a paid API. See our commercial interest.
What we collect
One row per published price for text and chat models. Each row holds the provider, the model, the provider's model ID where it shows one, the service tier, the input, cached-input and output price in USD per 1M tokens, the context window and numeric precision where stated, notes on region, tier or promotion, the source URL, the date checked and a verbatim quote from the page.
Today that is 1,671 rows from 41 providers (16 of them model makers with their own API), covering 430 models. 1,671 of 1,671 rows carry a quote, from 363 distinct source pages. Only 24.1% of rows state the precision served and 47.6% state a context window, because many pricing pages do not.
Where the prices come from
- Providers' own pages only. Each price is read from the provider's own pricing page or price documentation, never from a router, reseller or aggregator feed. Each provider has one data file in the site's code (
src/data/own-prices/). - Archive captures for history. The current prices are read from live pages. The separate LLM API price history reads past list prices from Internet Archive captures of the same providers' pricing pages: 3,337 rows from 11 providers since 2023-01-03, each with its archive URL and the exact text it was read from (how it was built).
How often prices are checked
- Check dates: every row has its own check date. Current rows were checked 2026-11-14 to 2026-11-15.
- Weekly re-check: a script re-fetches each source page and looks for the row's quote or prices. It flags what it cannot find and never edits data; a flag means the page must be re-read, not that a price changed. Flagged pages are re-read and the value, quote and date updated. Latest run 2026-11-16: 1,657 of 1,675 prices still matched, 18 are on pages that load prices with JavaScript and were read with a page renderer, and 0 prices were flagged for a re-read.
- Change log: the price change log is built from the history of the price files. A fix to our own data (a value first recorded, or any difference within 24 hours of a provider file's first version) is listed as a correction, not as a provider's price change.
- Snapshots: provider price database version 2026-11-15; archived price history built 2026-11-15; BroadNet rate card synced 2026-11-14.
How models are matched across providers
- Every row carries a model slug, and rows with the same slug are the same model on the model's price page. Providers' own model IDs are kept per row.
- The page name is the maker's own name for the model where it reads as a name, otherwise the name most providers use.
- A dated snapshot keeps its own slug. Where a maker serves a dated version under a base model name (2 cases today), the maker's row is shown on the base model's page, labelled as the maker's serving version, and is not ranked against hosts that sell a different snapshot.
- A row counts as the maker's own price only when the provider trains that model. A maker that also resells other makers' models (Alibaba Cloud sells DeepSeek and Kimi models, for example) is a host for those rows.
Which tier and region we use
- Standard tier ranks. Only standard, on-demand prices are ranked. Batch, Flex / spot, Fast, Ultrafast and Priority prices are separate rows, shown with a tag and never ranked.
- Region and prompt size: the price the page shows at the lowest prompt-size tier in the provider's default region. The row's notes name the region or endpoint where the page states one.
- Ranking key: a blend of 3 input tokens to 1 output token, (3 × input + output) ÷ 4. On a tie the maker's own API is listed first, then the lower input price.
- Promotions rank as published, with a "promo" tag and the end date or list price in the notes where the page gives one.
- Peak pricing: where a provider bills peak and off-peak rates (DeepSeek), the row carries the peak price and the off-peak rate is shown next to it.
- Free rows (zero input, cached and output price) are listed as "free" and never ranked, because they usually come with low rate limits or a time limit.
What we exclude
- Router, reseller and aggregator price feeds.
- Image, audio and embedding models. Speech-to-text prices are a separate dataset.
- Models with no public per-token price (for example "contact sales"); the provider file's notes say why.
- From search indexes: a model page is indexed only when 2 or more providers publish a sourced standard price for it, at least one of them paid. BroadNet's row never counts toward that. 198 of 428 model pages meet it; the other 230 are in a compact list and built with noindex. A provider page is indexed with 3 or more models, or 2 that another provider also prices.
How BroadNet's own prices are handled
BroadRouter is BroadNet's own site, so its prices are kept apart from the third-party data:
- No BroadNet price is typed by hand. BroadNet's prices come only from its published price file, pricing.json, synced 2026-11-14, and the build fails if a page contains a hard-coded dollar amount (policy).
- BroadNet is not one of the 41 providers. Its row on a model page is labelled "this site's owner", is never ranked and never counts toward the provider minimum.
- BroadNet is not in the provider price database or the price index.
- Where another provider's standard price is lower than BroadNet's, the model page says how many are lower.
What we do not measure
- Speed: no latency, time-to-first-token or throughput tests.
- Quality: no benchmarks or output-quality scores. The price ranking page applies a stated filter of its own; the model pages rank on price alone.
- Uptime or reliability: no availability monitoring.
- Usage: no traffic, market-share or popularity data.
- What people pay: these are published list prices, not invoices, negotiated rates or credits.
- Unstated facts: context and precision are shown only where the provider states them; "not stated" means unknown, not full precision.
AI assistance
Prices are read from providers' pages with AI assistance, and with a page renderer for pages that load prices with JavaScript. No named human reviewer signs off on each row, so a row can be wrong. The verbatim quote and source link on every row are there so you can check any price yourself, and the weekly re-check is automated, not a human review.
Corrections
Every model and pricing page has a link to report a wrong price, or email the page address and source as the editorial policy describes. Confirmed errors are fixed, the row's check date is updated and the change is shown in the price change log. Factual corrections across the site are recorded in the public fact-check log (60 entries, latest 2026-11-15), including ones that went against BroadNet's interest.
How to cite the price data
- One model's prices: cite its page under LLM API pricing with the check date shown on it, for example: "BroadRouter LLM API price tracker, prices checked 2026-11-15".
- The full data: BroadRouter (2026). LLM API provider price database: every price with its source, date and quote (version 2026-11-15) [Data set]. BroadNet Technologies. https://broadrouter.com/data/llm-api-provider-prices/
- Licence: Our compilation (the selection, normalised fields, model matching, notes, tables and method): CC BY 4.0. Each price is a fact published by the provider named in its row, at the source URL given per row, and each quote is a short excerpt from that page; we do not license those pages to you. Credit the provider as the source of each price.
Datasets built on these rows: LLM API provider price database (every row with its source, date and quote), LLM API price index (per-model statistics), LLM API price history (archived list prices) and the rest of our open datasets.
Questions
Where do BroadRouter's LLM API prices come from?
From each provider's own pricing page or price documentation: 1,671 prices from 41 providers, checked 2026-11-14 to 2026-11-15. No router, reseller or aggregator feed is used. Every row keeps its source URL, the date checked and a verbatim quote, and all rows are in the provider price database.
How often are the prices checked?
A script re-fetches every source page each week and flags rows whose quote or price it can no longer find; the latest run was 2026-11-16. It never edits data. Flagged pages are re-read and the value, quote and date updated, and the change appears in the price change log. Each row keeps its own check date.
Does BroadRouter rank BroadNet's own prices?
No. BroadRouter is BroadNet's own site. BroadNet's prices come only from its published price file and are shown in a separate, labelled row on model pages. They never count as a provider, never set a ranking and are not in the provider price database.
Does the tracker measure speed, quality or uptime?
No. It records published list prices and the context, precision and model IDs that providers state. It does not test latency, throughput, output quality, uptime or usage, and it does not record what any customer actually paid.
How do I cite the price data?
Cite the model page or dataset you used with its check date, and credit the provider named in each row as the source of that price. Our compilation is CC BY 4.0; a ready citation and BibTeX are in the provider price database.
Report a wrong price or number (email to info@broadnet.ai, with this page's link filled in).