Sources and method
255 models, 1,160 priced routes, 23 ways to buy and 311 dated changes so far: where each number comes from, and which companies are covered.
Coverage
The scope is every company that publishes a public per-unit price for access to AI models: labs, clouds, data platforms, hosts and gateways. Every company in the companies.fru.dev registry is enrolled and a few are checked each morning on their own pricing pages; the rest are checked by hand in research passes.
Bellwether companies: 20 of 20 in scope covered
10 out of scope| Company | Status | Where, or why not |
|---|---|---|
 Amazon | Covered | Amazon Nova |
| Covered | Anthropic | |
| Covered | Cloudflare Workers AICloudflare AI Gateway | |
| Covered | Cohere | |
| Covered | Cursor | |
| Covered | Databricks Mosaic AI Gateway + Foundation Model APIsDatabricks Mosaic AI | |
| Covered | Elastic Inference Service | |
| Covered | Google Gemini API | |
| Covered | Hugging FaceHugging Face Inference | |
| Covered | IBM watsonx.ai | |
| Covered | Meta | |
| Covered | Azure AI Foundry | |
| Covered | Mistral AI | |
| Covered | Voyage AI by MongoDB | |
| Covered | NVIDIANVIDIA NIM | |
| Covered | OpenAI | |
| Covered | OCI Generative AI | |
| Covered | Perplexity | |
| Covered | Snowflake Cortex AI (AI_COMPLETE)Snowflake Cortex AI | |
| Covered | xAI | |
No public per-token price. Foundation Models run on device; Private Cloud Compute access is free for apps in the App Store Small Business Program, and no other rate is published. Source | Out of scope | No public per-token price. Foundation Models run on device; Private Cloud Compute access is free for apps in the App Store Small Business Program, and no other rate is published. Source |
Rovo is priced in credits inside Jira and Confluence ($0.01 per credit past the allowance); no model API. Source | Out of scope | Rovo is priced in credits inside Jira and Confluence ($0.01 per credit past the allowance); no model API. Source |
Flink model inference connects to remote models on the customer's own provider; Confluent bills compute, not tokens. Source | Out of scope | Flink model inference connects to remote models on the customer's own provider; Confluent bills compute, not tokens. Source |
Sells observability, including LLM and agent monitoring and Bits AI credits ($500 per 500 credits); it does not sell access to models. Source | Out of scope | Sells observability, including LLM and agent monitoring and Bits AI credits ($500 per 500 credits); it does not sell access to models. Source |
GitHub Models, its model API, has been retired and is no longer available to any customer; Copilot bills premium requests inside the product, not tokens. Source | Out of scope | GitHub Models, its model API, has been retired and is no longer available to any customer; Copilot bills premium requests inside the product, not tokens. Source |
AIP is sold by contract; no public price list. Source | Out of scope | AIP is sold by contract; no public price list. Source |
Agentforce is priced per action, conversation and user (Flex Credits at $500 per 100k credits, $2 per conversation); no per-token price for model access is published. Source | Out of scope | Agentforce is priced per action, conversation and user (Flex Credits at $500 per 100k credits, $2 per conversation); no per-token price for model access is published. Source |
No public price list; sales-led ("Book a 1:1 intro call"). Source | Out of scope | No public price list; sales-led ("Book a 1:1 intro call"). Source |
No public price list for Now Assist; the pricing page answered 403 to our reader and was not retried. Source | Out of scope | No public price list for Now Assist; the pricing page answered 403 to our reader and was not retried. Source |
Sells payments and usage-based billing for AI companies, not model access. Source | Out of scope | Sells payments and usage-based billing for AI companies, not model access. Source |
46 official prices read by hand from companies' own pages so far. The bellwethers are the family's shared audit list.
Model prices
Every morning at 05:00 UTC the site reads OpenRouter's public model catalog (openrouter.ai/api/v1/models) and, for each tracked model, its endpoint list, which carries the price each provider charges. The lab's own endpoint is the list price; Bedrock, Azure, Vertex AI and hosts are separate routes, with regions kept apart. Batch prices come from the catalog's batch variants. Prices are USD per million tokens.
History
Before daily tracking began, history comes from the Internet Archive's snapshots of the same catalog (about 190 captures since March 2024). A change is dated by the first snapshot that shows the new price, and links to it. Only closed-weight models are backfilled, because an open model's catalog price is simply the cheapest host that day.
Gateways
Markup, bring-your-own-key fees, free tiers, plans and features were read by hand from each gateway's own pricing page and docs, with the quotes kept on each gateway page. Every Monday the pages are fetched again; a page whose text changed is re-read by a language model and the result waits, marked Unverified, until it is checked by hand.
A gateway is priced only on models it sells. Every morning each gateway's own model catalog is read (a JSON feed where it publishes one, such as OpenRouter, Vercel, Neon, Helicone or Requesty; the providers page for bring-your-own-key proxies like Portkey and Kong; the official model list for clouds and data platforms, such as Bedrock's model cards, the Vertex AI pricing page, Oracle's price list, the watsonx model specs and the Snowflake and Databricks docs, plus any model priced here on that platform) and matched to the models tracked here. Official models that match nothing tracked here are kept as leads for the catalog. Where a gateway does not list a model, the cost tables show a dash (Not offered) instead of a price. A model that leaves a list is kept with the date it left, and a list that fails to load or suddenly halves is left as it was.
Official price lists
Some companies sell model access the catalog does not carry: embeddings and rerankers, search APIs, platform credits, GPU hours and licences. Their prices are read by hand from the company's own pricing page or docs, copied as the page states them (nothing converted or estimated), and each row links to the page it came from. Later passes re-read them and every change is kept. Companies checked and found out of scope keep the reason and the page that shows it, and are checked again every few months.
Quality scores
The quality-per-dollar overlay uses the overall score from leadersboard.fru.dev, divided by the blended price (3 input tokens to 1 output).
Disclaimer
LLM API list prices from OpenRouter's public catalog and each provider's own pricing page, read every morning, plus price lists read by hand from companies' own pages, each linked to its source; gateway fees are re-read every Monday and changes wait, marked unverified, until checked by hand. The list grows as providers launch new models. Your contract may differ; not investment advice. Logos via logo.dev; trademarks belong to their owners. Anonymous usage stats via Google Analytics. Volume discounts, committed-use deals, regional pricing and taxes are not included.
Corrections
Every model and gateway page has a Suggest a correction link. Each suggestion is checked against its source before anything changes.