Compare costs

Closed or open?
Compare your OpenAI and Anthropic costs with open-weight models in Europe.

With your own numbers, no account, in about a minute.

Most teams pay OpenAI or Anthropic per token or per seat. Open-weight models such as DeepSeek, Kimi and GLM do much of the same work for a fraction of that price, hosted in Europe. Below you check that against your own usage: what you pay now, what the same tokens cost here, and what a GPU of your own would do.

Try it with your own numbers

You need no account and we store nothing. Pick whatever you have at hand: an admin key (we read your tokens of the last 30 days once and discard the key), your tokens or monthly spend per model, or the subscription you have today.

Where do I find my admin key?

OpenAI: platform.openai.com, Settings (gear), Organization, Admin keys, Create admin key. You need to be an Owner of the organisation. The key starts with sk-admin and is shown once.

Anthropic: console.anthropic.com, Settings, Admin keys, Create Admin API Key. You need to be an Admin of the organisation. The key starts with sk-ant-admin. Regular keys under API keys (sk-ant-api) do not work.

An admin key can do more than read. Delete it on that same page after the check.

The same maths as our billing. Your side: list price or seat price. Our side: what you pay here, margin included.
Read access to usage only Free, no account needed Key is not stored
−75%
for a startup of 8 on ChatGPT Team

The same traffic on open weights through the router, at today's list prices. Other teams are listed below.

75
OpenAI and Anthropic models with a list price

Refreshed daily, converted to euro. Every model gets an open-weight counterpart next to it.

0
keys stored, accounts needed

One read, report shown, key gone. We keep neither your tokens nor your subscription.

Three forms

The same open-weight models, three ways to run them

Your app keeps its OpenAI or Anthropic SDK; only the base URL changes. Behind it you choose per application: shared per token, a dedicated GPU with us, or your own hardware.

Your app
OpenAI or Anthropic SDK, unchanged
HostYourAI Router
EU, one base URL
Shared, per token
warm open-weight models
Dedicated GPU
your own machine, VPN, per hour
Self-hosted
your rack, the same models
drop-in warm · EU
The report prices your traffic on the first two; self-hosting runs the same models on your own hardware.
Examples

What teams save when they move from ChatGPT or Claude to open-weight models

Priced with today's list prices and our own rates. Click a case and the form above fills itself; you then get the full report, with every model side by side.

When

When open-weight models beat OpenAI and Anthropic on cost, and when they do not

From what we see with customers and in our own numbers. No rule is absolute, which is why the check above exists.

High volume, repetitive work

Support, classification, extraction, summaries, translation. This is where the gap is biggest: usually 70 to 95 percent less, and you rarely notice a difference in quality. This is where almost everyone starts.

Coding agents with a lot of context

Agents resend the whole conversation every turn, and Anthropic charges little for those cached tokens. The gap is then smaller than you would expect. Open weights win here with a mid-size model such as DeepSeek V4 Pro or GLM 5.2, not necessarily the very largest.

Teams on seats

A team plan costs per person, whether that person uses a lot or a little. Per token you only pay for what actually happens, and for most teams that sits far below the seat price. It also saves the discussion about who gets a seat.

Personal data or regulated work

Customer records, HR files, medical intake, insurance claims. On a GPU of your own in the EU with a VPN, your data stays on one machine, no third party sees a prompt and nothing is trained on it. With a processor agreement in place you can process personal data there; the legal basis stays yours, as with any processor.

Example: a claims workflow of 300 million tokens a month on DeepSeek V4 Flash costs €92.40 a month through the router with the EU switch, and €8,753 a month on a dedicated 4x RTX PRO 6000 (392 GB), fully isolated. Which one you need is a compliance choice, not a price choice. More on security

When you are better off staying

If you rely on one specific strength of a closed frontier model, or if your usage is small. Then the difference is a few euros a month and switching costs more time than it saves. You will see that here too.

When a GPU of your own pays off on price

From billions of tokens a month on one model. Below that, per token through the router is almost always cheaper, because a machine of your own keeps running at night and at the weekend. The calculator on the pricing page shows the break-even point. To the calculator

How it works

How the cost comparison works

You change nothing in your current setup. The check only looks.

01Give us your numbers

An OpenAI or Anthropic admin key, your tokens or monthly spend per model, or the subscription you have today. Whatever you have at hand.

02We price the same traffic on open weights

Every model you use gets a comparable open-weight model next to it, per token through our EU router and, when that comes within reach, on a GPU of your own.

03You decide with the report in hand

Per model what it costs now and what it costs here, plus the same tokens on the popular coding models. Print or save as PDF straight away.

Counters only. With a key we read your provider's usage report: tokens per model per day, nothing else.
Nothing stored. Key used once and discarded; we keep neither tokens nor subscription.
The same prices as the invoice. Our side uses the billing columns, margin included.
Europe. Every provider behind the router is on the sub-processors page; with the EU switch processing stays with parties established in the EU.
FAQ

Questions about moving from OpenAI or Anthropic to open weights

How much cheaper are open-weight models than OpenAI or Anthropic?

That depends on your workload. For support, classification, extraction and summaries we usually see 70 to 95 percent less. For coding agents with a lot of context the gap is smaller, because Anthropic charges little for cached tokens. For teams on seats the gap is often the biggest, because per token you only pay for what actually happens. The check above works it out with your own usage.

Which open-weight model replaces GPT-5, Claude Sonnet or Claude Opus?

We put GPT-5 next to DeepSeek V4 Pro, Claude Sonnet next to GLM 5.2, Claude Opus next to Kimi K3 and the small models (GPT-5 mini, Claude Haiku) next to DeepSeek V4 Flash. That is a suggestion based on what customers pick in practice, not a benchmark. The report lists every popular coding model side by side with your tokens, so you see at once what a step up or down costs.

What do I need to compare: a key, my tokens or my subscription?

Whatever you have at hand. With an OpenAI or Anthropic admin key we read your real usage of the last 30 days. If you know your tokens or monthly spend per model, you enter those. If you pay for a ChatGPT or Claude subscription, you pick the plan and the number of users. All three give the same report.

What exactly do you read with an admin key, and is it safe?

Only your provider's usage report: tokens per model per day. No prompts, no answers, no end-user names. The key is used once, never for a request to a model, and then discarded. Delete it on your provider's side after the check as well; then the loop is closed.

Where do I find my OpenAI or Anthropic admin key?

OpenAI: platform.openai.com, Settings, Organization, Admin keys, Create admin key; you need to be an Owner of the organisation, the key starts with sk-admin. Anthropic: console.anthropic.com, Settings, Admin keys, Create Admin API Key; you need to be an Admin, the key starts with sk-ant-admin. A regular API key cannot read usage.

Is the comparison fair?

Your side uses the public list price or seat price, converted to euro; batch discounts or negotiated rates are not included. Our side uses exactly the prices you are billed at here, margin included. Cached tokens count at the cache rate on both sides. Where we are more expensive, it says so.

When is a dedicated GPU in the EU worth it?

On price only from billions of tokens a month on one model; below that, per token through the router wins, because a machine of your own keeps running at night. On compliance sooner: whoever processes personal data or regulated work and wants no third party in the chain at all picks a dedicated GPU with a VPN. The report shows a dedicated GPU as soon as it comes close to what you pay now.

Can I process personal data (GDPR) on open-weight models?

Yes, with a processor agreement, as with any processor. Through the router with the EU switch, processing stays with parties established in the EU, all listed on our sub-processors page. On a dedicated GPU with a VPN your data stays on one machine and nobody else sees a prompt. In both cases nothing is trained on your data. The legal basis for the processing stays yours.

Do I need to be a customer already?

No. You compare right here, without an account. You only switch once the numbers convince you, and it does not have to happen all at once: the router speaks both the OpenAI and the Anthropic API, so you can switch one application at a time.

Check it with your own numbers.

It takes a minute and we keep nothing. Switching is a separate decision after that, one you make with the numbers in hand. Not sure about a model? We are happy to look with you.