AI Model Prices: Cost, Speed and Uptime by Provider

What large language models actually cost to run, in money rather than scientific notation, and which company hosting each one is cheapest, fastest and most reliable right now.

Hundreds of models across every major vendor. No API key, no account, no browser.

Modes

• Models - one row per model: price per million tokens in and out, cache read and write rates, the cost of a realistic request, context window, modalities, tool and reasoning support, knowledge cutoff and retirement date. Cheapest first.
• Providers - one row per model per hosting company: its own price, measured latency and throughput, uptime over 5 minutes, 30 minutes and a day, quantisation and maximum output.
• Compare - one row per model naming the cheapest, fastest and most reliable provider, the price spread between best and worst, and any provider currently degraded.

Example output

{
  "mode": "models",
  "modelId": "openai/gpt-4.1-nano",
  "inputPricePerMillionTokens": 0.1,
  "outputPricePerMillionTokens": 0.4,
  "requestSizeInputTokens": 10000,
  "requestSizeOutputTokens": 1000,
  "costOfRequestUsd": 0.0014,
  "costPer1000RequestsUsd": 1.4,
  "contextLengthTokens": 1047576,
  "hasTieredPricing": false,
  "costRankCheapestFirst": 7
}

The two things a price list will get wrong

1. Prices are published per token. A rate arrives as 0.00000003. Nobody reasons in those units, so everything here is converted to cost per million tokens and to the cost of a request you actually specify — set your own input and output sizes and get an answer in dollars.

2. Pricing is tiered by prompt length, and the headline rate hides it. One model in the catalogue charges $0.03 per million below 32,000 tokens, $0.10 above it, and $0.20 above 256,000. Costed at a 300,000 token prompt, the real bill is 6.7 times what the headline rate implies. The tier that applies to your request size is the one used, the threshold is named in tierAppliedToThisRequest, and every tier ships in pricingTiers.

Other things worth knowing

• A missing rate is not a free one. Only 27 of the models publish an image price and 211 publish a cache read rate. Where a usage type is not priced it stays null and is named in unpricedUsageTypes, rather than collapsing to zero and implying that usage is free.
• Speed is not always measured. Latency and throughput come back empty for providers the aggregator has not sampled, and Number(null) is zero — which would publish a working provider as "0 tokens per second". Unmeasured providers report null with speedMeasured: false.
• The same company can host one model more than once, at different prices, regions or reliability. Those endpoints are ranked individually rather than collapsed by name, so a cheap endpoint and an expensive one from the same provider both show honestly.
• Degraded providers are flagged. A non-zero status from the aggregator sets isDegraded, and compare mode lists them.
• Prices move. This is a feed to schedule, not a fact to cache. It also reflects one aggregator's catalogue and its listed rates, which can differ from a provider's own direct pricing — stated on every row.
• Provider detail costs one request per model, so those modes expand the cheapest matching models first, up to a limit you set.

Who this is for

Engineering and product teams choosing a model or a provider, finance and analytics teams modelling inference spend, anyone routing traffic who needs to know which endpoint is cheap and actually up, and analysts tracking where model pricing is heading.

Pricing

$0.004 per row. The first 2 rows of every run are free, and note rows (a filter nothing matched, a model with no provider detail) are never charged.

A vendor's full catalogue is usually 20 to 70 rows. Comparing providers across 15 models is 15 rows in compare mode, or roughly 60 to 100 in providers mode.

Related actors

• Hugging Face AI Models Scraper - model metadata, downloads and popularity, which this deliberately does not duplicate.
• Package Adoption Tracker - npm and PyPI download trends for the tooling around these models.

How to run it via API

curl -X POST "https://api.apify.com/v2/acts/scrapemint~ai-model-prices/runs?token=YOUR_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"mode":"compare","vendors":["anthropic","openai"],"promptTokens":10000,"completionTokens":1000}'

Catalogue and provider measurements from OpenRouter.