AI models and provider keys
Every session, agent run and evaluation in FlowLakes calls one model. This page is about which models exist, whose key pays for the call, and how a tenant changes either.
Platform-provided models
FlowLakes ships a list of platform-provided models. They appear on the Connections page, under the model card, from the first minute of a new tenant. Each one names a provider and the key it uses:
| Model | Provider | Key held by |
|---|---|---|
| Claude Opus 5.5 (the default), Claude Opus 4.6, Claude Opus 4.7, Claude Opus 4.8, Claude Opus 5, Claude Sonnet 4.6, Claude Sonnet 5, Claude Haiku 4.5 | Anthropic | the operator |
| GPT-4 family | OpenAI | the operator |
| Mistral Medium 3.5 | Mistral, hosted in the EU | the operator |
| Mistral Large 3 (needs a Mistral tier that carries the Large models) | Mistral, hosted in the EU | the operator |
| Mistral Large 4 (public preview, same tier condition) | Mistral, hosted in the EU | the operator |
| Text Embedding 3 Small (memory index only) | OpenAI | the operator |
“The operator” means the organisation running your instance. The key is a setting of the instance, never a value stored in your tenant, never shown in the interface, and the same key serves every tenant on that instance. Whether platform AI use is part of your plan is set out in your agreement with FlowLakes.
Mistral Large 3 is pinned to its dated version. FlowLakes does not use a provider’s rolling “latest” alias, because that would change the model under your tenant without a decision. Mistral Large 4 is a public preview and has no dated version yet; its row says preview, and it will be pinned when Mistral publishes a dated id.
The default model
One platform-provided model is the default: Claude Opus 5.5. A new tenant’s sessions use it until an administrator chooses another one on the model card. The choice is tenant-wide, needs the model write capability, and is stored once, on the tenant’s own configuration. Every session and agent run reads that one choice first; the seeded platform default only answers for a tenant that never chose.
The embedding model is separate. It only turns the tenant’s operational memory into vectors for recall, and it cannot be switched from the model card: the memory index stores vectors of one fixed size.
The monthly AI budget
No plan caps AI use. What a tenant spends is metered and shown on the AI settings page, under What it has used, for today, this week and this month, in billable tokens and in estimated money. Money is an estimate from the per-token prices on the model rows. The provider invoice is the record.
An administrator can set a budget of the tenant’s own under Monthly AI budget: a number of billable tokens a month, an estimated spend in dollars a month, or both. Leave a box empty for no limit. A budget is checked against this calendar month’s spend on every AI call. Once a budget is used up, every AI call is refused until the month rolls over or the budget is raised, and the person who hit it reads which budget it was and the numbers.
For each budget that is set, the card shows this month’s spend, what is left and the percent used. It turns amber at 80 percent and red when the budget is used up. With no budget set, the card still shows this month’s spend as plain information.
A fresh tenant has no budget. Nothing is seeded on its behalf.
Bringing your own key
A tenant that wants its own provider account sets it on the model card, under Connections: a single key field next to the model in use. Saving it creates a tenant-owned copy of that model carrying the key, encrypted at rest and never returned by any endpoint or shown again once saved. From that point the resolver prefers the tenant’s own row over the platform-provided one for every session and agent run in that tenant, and the operator’s key is no longer used there. Removing the key on the same card falls the tenant back to the platform-provided model; nothing else on the tenant’s configuration changes.
What leaves the tenant for the provider
Every model call goes through the privacy shield at the tenant’s masking tier before it leaves. Which values are masked, and which are not, is described in Privacy and masking. The memory index is embedded under the same shield.
Platform model keys
On hosted FlowLakes, platform model keys are managed by FlowLakes. A platform-provided model whose key is not set refuses with a typed, fixed error; it never falls back to another provider and never shows the tenant the raw provider or key name. The model card on Connections reads the same answer: a model whose key never resolves shows as not configured there, the same way for every tenant.