Explore Flowstate

The Flowstate platform
Flowstate Helm

Know what AI costs. See what you get back.

Understand spending, share budgets and connect AI use to the work it supports.

Recent insights All insights

Buy tokens, rent GPUs, or own the rack?

Open-weight models are now good enough to make self-hosting a financial question rather than a capability one. I modelled four options, an office rack, colocation, rented GPUs and per-token APIs, to find where each one actually wins. The answer turns on utilisation, and it isn't the one the internet keeps giving.

Turn a $3m AI bill into $1.9m

Most teams overpay for AI twice without noticing: they run everything through the default flagship model, and they buy the tokens on a metered plan they can't see into. An intelligent proxy fixes both. Here is the maths, a calculator you can drag, and where Flowstate fits.

Start with the work All use cases

Understand AI spending and value

See where the budget goes and what the business gets back.

Share one AI budget across your tools

Fund people, teams and projects across connected services.

Give your team more capacity

Reduce the preparation, checking and follow-up around recurring work.

Connect AI use to delivered work

Follow AI activity into projects, code and business results.
Quick links
Flowstate / People + AI Talk through your priorities
Documentation
Get help
Docs / Flowstate Router

How routing works

By default, Flowstate Router passes each request to the model the tool asked for. Two optional features can change that: intelligent routing and outage failover. An admin turns each on. Both are off until then.

Intelligent routing

Intelligent routing can move a simple prompt to a cheaper model from the same provider and family. A quick question doesn’t need the most capable model.

  • It’s set per workspace, separately for people and for agents.
  • It only moves a request to a cheaper model in the same family. It doesn’t switch provider.
  • It looks at what the prompt asks for. It leaves complex work on the model the tool requested.

Decisions are made per thread

Flowstate decides once per conversation thread, not once per request. Then it keeps that choice while the provider’s prompt cache is warm.

This matters because AI providers keep recently used context in a cache, and a cached prompt costs less to reuse. Switching models in the middle of a thread would throw that cache away. Keeping the decision stable means routing doesn’t make a conversation more expensive by discarding cached context.

If a later turn needs a more capable model, Flowstate moves the thread back up.

Outage failover

If an AI service isn’t responding, Flowstate can send the request to a Flowstate-hosted model instead, so people can keep working.

  • It needs your explicit consent. Flowstate doesn’t send requests to its own models until you give it.
  • It’s for outages. It isn’t a way to choose cheaper models.

Failover models are different from the model the tool asked for. Decide whether that’s acceptable for your work before you consent.

Every decision is recorded

Flowstate records what happened to each request:

  • The model the tool asked for.
  • The model that answered.
  • Whether the request was moved to a cheaper model, and why.
  • Whether it failed over.

These records are kept for every request, whether or not session transcripts are on. They hold no conversation content. Admins use them to check that routing behaves as expected.

What routing doesn’t do

  • It doesn’t guarantee savings. What you save depends on how people use AI.
  • It doesn’t change the AI account people sign in with.
  • It doesn’t route when it can’t recognise the person. Those requests are refused.