Everyday requests to a cheap model.Demanding ones to a big one.

See current prices here.

The reference price is the public list price per million tokens of the most expensive model you allow. It is preset for you, so getting started requires nothing, and you can change it in the portal. The saving is the difference between the reference price and the actual price for exactly the same requests. The request log shows both prices for every request, and you can download the log as a CSV file and calculate it yourself. If a period ends up negative, the report shows it as it is, and no savings fee is charged.

A cheaper model’s answer is returned only if it passes the quality gate. Otherwise the request moves up to a bigger model. On top of that, a sample of requests also runs on the strongest model your rules allow, and the answers are compared side by side. The result appears in the pilot report. For a task where you want no compromises, you can lock the model. Then no savings arise, and none are billed either.

You don’t have to trust blindly. Create a separate key for AIvantage in your model provider’s console. You can see in the provider’s own view what it is used for, and revoke it there in seconds. AIvantage stores the key encrypted (AES-256-GCM) and never shows it to anyone after it is saved. From prompts and responses we store only training excerpts for your own savings model. They are deleted after 180 days, and collection can be switched off.

Savings apply to traffic through the API, meaning applications, integrations and coding assistants that use an API key. AIvantage does not connect to per-user licences such as ChatGPT or Copilot subscriptions. If there is no API traffic, there is nothing to save, and no invoice either.

In an OpenAI- or Anthropic-compatible application, including Claude Code, getting started begins with changing the address and key. Automatic routing additionally requires the model to be set to auto or aivantage/auto. The embeddings API, Batch and Files API are not supported. We check the features your application needs at the start of the pilot.

If the model provider doesn’t respond, the request moves to the next model your rules allow. If AIvantage itself doesn’t respond, your application gets an error. You then put the model provider’s own address and key back in the application, and traffic flows directly again. In our own production use in October, uptime was 99.67%, against an agreed level of 95.5%.

A little. Shield adds under a millisecond to a typical request. Name detection takes more, about 65 ms per thousand new characters, when names are blocked. When the quality gate rejects a cheaper model’s answer, the query goes to a bigger model, and that request takes longer. In the pilot, latency is measured on your own traffic.

From the invoices. The report shows in euros the usage of every use case and key. If the model provider’s own invoice is larger than the usage that passed through AIvantage, the difference tells you that someone is using AI outside your policy.

You get a summary with the savings in euros and percent, Shield’s findings and the result of the quality spot checks. Then you decide whether to continue. If you continue, Savings costs 20% of the savings actually achieved. If you don’t activate the service, the API stops when the pilot ends, and nothing is billed for the pilot. You get a summary of the pilot savings by email.

Only your own savings model. Your data trains only your own savings model – never shared models, other customers or our own purposes. Training excerpts are deleted after 180 days, collection can be switched off, and data is deleted on request.

You switch the application’s address and key back to the model provider’s own. Then you remove your key from the portal and revoke the AIvantage keys. It’s worth downloading the logs and reports as CSV files before you leave.