Talos AI tokens

An AI allowance on hosted Talos, for organizations without their own provider key.

Talos’s AI features need a model. You can always bring your own provider key. On hosted Talos, an organization can also claim Talos AI tokens: an allowance of AI usage that runs on Talos’s own model account, so the AI features work before you have set up a key of your own.

Hosted Talos only. A self-hosted install has no Talos AI tokens: it uses the provider keys you add under Settings → AI.

Claiming

An Organization Admin or Admin opens Settings → AI → Talos AI tokens, adds an optional note about what you plan to use Talos for, and claims. Every claim is reviewed and approved by hand, so it shows as pending until then. An organization has at most one pending claim, and cannot claim again while an allowance is waiting or in use. A declined claim shows the reason, and you can claim again straight away.

How much, and for how long

An approved allowance is 1,000,000 tokens. The 30 days start with your organization’s first AI request paid with them, not at approval, so nothing is lost while you get set up. After 30 days the allowance expires whether or not tokens remain. An approval that is never used lapses 30 days after it was approved. Either way you can claim again.

Input and output tokens count as the model provider reports them. Before each request, Talos sets aside the most that request could use; the last request before the limit can go over it by at most 20,000 tokens. At most 3 requests per organization run on the allowance at once — more wait briefly.

Which models

Requests paid with Talos AI tokens use OpenAI models only: gpt-6.1-terra for planning (agent planning and failure diagnosis), gpt-6-luna for everything else, and text-embedding-3-small for search indexing. Your own key always comes first: the allowance pays only for features with no model of your own, unless an admin chooses otherwise per feature in Settings → AI → Routing.

Sharing it fairly

Admins can give a person a cap — a share of the allowance in tokens — and cap background AI jobs at a percentage of the allowance, in the Limits section of the same page. A person who reaches their cap is told so; everyone else is unaffected. Two requests at the same moment can take someone over their cap by one request; the organization’s allowance itself is the hard limit.

Background AI jobs — work the AI starts on its own, such as diagnosing failed jobs or a nightly project digest — are switched off for a new hosted organization. Turn each one on in Settings → AI → Background jobs; turning one on first shows an estimate of the tokens it will use and who pays for them. A background job that is skipped, because it is off or a cap was reached, is recorded with the reason.

Emails

Talos emails your organization’s owner and admins when a claim is approved or declined, when 80% of the allowance is used, when it is all used, 3 days before it expires, and when it has expired or an unused approval has lapsed. Each is sent at most once per allowance. When the allowance is used up, AI requests paid with it stop with a clear message; add your own provider key to carry on.

Privacy

Requests paid with Talos AI tokens are sent to OpenAI under Talos’s account, including the infrastructure context the AI needs to answer: OpenAI processes that data for Talos. Your AI policy’s data-egress setting still applies — with redacted only, the default, context is redacted before it leaves Talos. For the allowance, Talos stores token counts, not your prompts or answers. If your organization’s AI policy does not allow OpenAI, Talos AI tokens are not used.