DEV Community

Cover image for Same AI model, five weeks later: the cheapest provider had changed
uehara
uehara

Posted on Edited on

Same AI model, five weeks later: the cheapest provider had changed

One AI model, DeepSeek V4 Flash, was on sale from 17 providers through the same relay service. For one job of 100,000 input and 1,000 output tokens, the cheapest provider was OpenInference. About five weeks later, OpenInference was gone from the list, and Wafer, the third cheapest, had raised its input price about 2.7 times.

The prices were checked by an AI worker in our company's AI office, and checked again for this post.

Key points

  • The same model costs different amounts depending on the provider. Across the 17 providers in the first check, input ran from $0.0138 to $0.44 per million tokens, a spread of over 30 times.
  • Prices change within weeks. About five weeks later, OpenInference, the cheapest, was gone; Wafer had raised its prices; and Relace, which had cut its prices, was now the cheapest.
  • Check the latest prices each time before you use a provider. A price you looked up earlier, or read in an article, may not be today's price.

Checking the price list

The trigger was a Japanese post on note by zephel01, "Chinese LLMs: which provider you go through changes your bill more than which model you pick — DeepSeek/GLM/Qwen/Kimi" (title translated). It said cache prices for the same model can differ tenfold depending on the provider.

Instead of reusing those numbers, the AI worker checked the same list itself. The source is OpenRouter, a relay that lets you reach one model through many providers. For each model, OpenRouter publishes which providers serve it and at what price, through an API you can read without logging in.

The worker read that list from Python and saved the response, a block of data in a format called JSON, as it came. It sent no jobs to the model itself, so the whole exercise cost nothing. DeepSeek V4 Flash had 17 providers.

The price fields the public OpenRouter API returned for OpenInference and Wafer. Prices are strings in US dollars per token: regular input (prompt), input read from cache (input_cache_read), output (completion) and discount. OpenInference charges the same for regular and cached input; Wafer charges less for cached input. The saved response, processed again with the same formula for this post.

Each price comes back as a string in US dollars per token, in three price fields: regular input, input read from cache, and output. There is also a discount field. The worker multiplied those strings as exact decimals (Python's Decimal) to avoid rounding errors.

Pricing one job

The worker priced one made-up job of 100,000 input tokens and 1,000 output tokens for each provider: input tokens × input price + output tokens × output price.

The 12 compared providers from the first check, sorted by cost with no cache. Columns: provider, input, cache and output price per million tokens in US dollars, then cost and rank with no cache, and cost and rank when 90% of the input is read from cache. Highlighted: OpenInference (first with no cache, $0.001917) and Wafer (third with no cache, $0.004670).

OpenInference was the cheapest at $0.001917 per job. Relace was second at $0.003010, and Wafer third at $0.004670. The most expensive, Cloudflare, cost $0.04532, more than 20 times OpenInference. The two columns on the right are the cost when caching applies, covered in a separate section below.

What was left out

Of the 17 providers, five were set aside and the remaining 12 were compared.

  • Mancer 2, whose cache price was blank. Treating a blank as zero would make it look unfairly cheap, so the blank stayed blank and Mancer 2 was left out.
  • Baidu, StreamLake, GMICloud and Venice, which had a non-zero value in the discount field. That field is a separate value applied to the prices, so these were left out to keep the conditions equal.

The same model name can also carry labels such as fp4 or fp8, which show how finely the provider stores the model's numbers. The answers are not guaranteed to be the same. After narrowing by price, test the answers on your own work.

Five weeks later, the prices had changed

Before writing this post, the worker checked the same list for the same model again through the same API. In about five weeks, several prices had changed.

The first check and the check about five weeks later, provider by provider, in US dollars per million tokens. Values that did not change show one number. DeepInfra offers several endpoints at different prices, so its rows are split by label (fp8 and so on). Alibaba, Cloudflare, DigitalOcean, Novita and OpenInference were gone; Fireworks and DeepInfra's flex and priority endpoints were new. Highlighted: OpenInference (gone), Wafer (input 0.0450 → 0.1200, cache 0.0130 → 0.0240, output unchanged at 0.1700) and Relace (input and cache 0.0173 → 0.0140, output 1.2800 → 0.4500).

  • OpenInference, the cheapest, was gone. Alibaba, Cloudflare, DigitalOcean and Novita were gone too, five in all.
  • Wafer, third, raised its prices. Input went from $0.045 to $0.12, about 2.7 times, and cache from $0.013 to $0.024.
  • Relace, second, cut its prices. Input went from $0.0173 to $0.014, and output from $1.28 to $0.45.
  • Three were added. Fireworks, and two more DeepInfra endpoints at different prices (flex and priority). That brought the count from 17 providers to 15.

Besides Wafer and Relace, Baidu, StreamLake and Mancer 2 also changed their prices, five in all. Those three went down, and Mancer 2 now lists a cache price where it had a blank.

Pricing the same job with the new prices gives this order.

The 11 compared providers from the check about five weeks later, sorted by cost with no cache. Highlighted: Relace (first, $0.001850) and Wafer (fifth, $0.012170).

Relace was now the cheapest at $0.001850, up from second at $0.003010. Wafer went from $0.004670 to $0.012170 and fell from third to fifth. OpenInference, the first-place provider, could no longer be picked at all.

The note post named DeepSeek's own service as the provider with the cheapest cache price. It was not in either of the two lists the worker checked. A price list from August cannot stand in for today's.

A separate point: caching changes the order again

This is a separate point from the price changes. Caching means the provider keeps a part you already sent and charges a lower price to read that part again in the next request. When you hand the model the same material many times, much of the input comes from cache.

Using the prices from the first check, the worker also priced the job with 90% of the input read from cache. The 90% is not measured; it is a placeholder to see how the cost moves. In that case Wafer was the cheapest at $0.001790, and OpenInference stayed at $0.001917 in second place. OpenInference charges the same for cached and regular input, so caching does not lower its cost. Wafer charges about 3.5 times less for cached input.

Cloudflare, the most expensive with no cache, drops to $0.00698 with 90% cached, about 6.5 times less. If you rule out providers on the input price alone, you can miss ones that get cheap for work where caching applies.

Price your own workload with the latest prices

AI prices differ by provider and change within weeks. A price you looked up earlier, or read in an article, may not be today's. This time, in about five weeks, the cheapest provider disappeared from the list.

Before you use a provider, do these three things each time:

  1. Check the latest prices in the public price list and save what comes back as it is.
  2. Price your own workload: its input size, output size, and how much of the input will be cached.
  3. Leave out blanks and fields whose meaning you have not checked. Do not treat them as zero.

An AI worker can do all three. What a person checks is whether the assumed workload and cache share match the real work, and whether the candidates' answers are good enough. These are estimates from public prices, not bills, and they leave out tax and speed.

Terms

  • Provider: a company that runs an AI model and serves it to users.
  • OpenRouter: a relay service that lets you reach one model through many providers.
  • Token: the unit AI models use to count text. OpenRouter returns prices per token; this post shows them per million tokens.
  • Cache: the provider keeps a part you already sent, and charges a separate (cache) price to read that part again in the next request.
  • fp4, fp8: how many bits the provider uses for the model's numbers. Fewer bits run lighter but can change the answers.
  • API: an interface programs use to read and write data.
  • JSON: a common text format programs use to pass data to each other.

Articles on our fight with cloud costs — attribution, measurement, structural guards, and continuous observation — will continue as a series. If you are interested, please like and follow.

Costwary, our cost-management service built on this methodology, is available at https://costwary.com
Our other products are listed at https://www.eln.ne.jp/products


EarthLink Network builds its own products to run the whole company on AI — each one born from a need along the way. See the full index of what we build:

→ The 18 products EarthLink Network builds in-house

For more about the company and each product, visit www.eln.ne.jp.

Top comments (0)