Models and the Wallet
Use SynapseOrch's platform models on a prepaid wallet, bring your own provider keys, or mix the two.
On SynapseOrch Cloud you have two ways to get a model, and you can use both at once.
| You pay | Set up | |
|---|---|---|
| Platform models | SynapseOrch, from a prepaid wallet | Nothing — they are already there |
| Your own keys | Your provider, directly | An admin adds keys, as in self-hosting |
Platform models
Available on every plan, including Free. There is no tier that unlocks them — what governs them is your wallet balance, not your subscription.
Owners and admins top the wallet up under Settings → Billing (Add Balance), and model usage draws against it as runs execute. The balance is authoritative at the moment of spend, so the figure you see is the figure you can actually spend — it is not an estimate that catches up later.
The wallet stops at zero
When the balance cannot cover a call, platform-model calls stop rather than continuing and invoicing you afterwards. This is the same shape as every plan's credit limit: a hard stop is predictable, and a surprise invoice is not.
A run the wallet stopped says so on its page — This run stopped because the model wallet ran out. — with Add balance, and the step names what was left and what the call needed. A chat message the wallet cannot cover fails; Try again sends it again once you have topped up.
There is no automatic top-up yet: Settings → Billing lists it as coming. Watch the balance, and top up by hand.
Every plan, Free included, comes with a $2.00 monthly model allowance of platform models, so you can try them without entering payment details. The allowance pays for any model we offer, before your balance, and what is left at the end of the month does not roll over. A balance you buy works on any model too.
Knowledge draws on the wallet too — embedding documents and questions, reranking, reading scanned pages and writing contexts — at the prices in Costs and budgets.
Bringing your own keys
An owner or admin adds provider keys under Settings → Models, exactly as you would when self-hosting, and those models become available to your agents. The model half of your bill is then zero — you pay OpenAI, Anthropic or AWS directly and SynapseOrch Cloud charges only for execution.
This is usually right when you already have committed spend or negotiated rates with a provider, or when your organisation requires the provider relationship to be yours.
Keys are held per organisation, read for each run, and never written into the process environment — the same boundary as self-hosting.
Mixing them
Per-agent and per-step model overrides mean the choice is not all-or-nothing. A common shape is your own key for the high-volume workhorse model, and platform models for the occasional expensive one you do not want a separate contract for.
What you were charged for
The model actually used is recorded in each run's trace. Where a model is selected automatically, that record is what makes the charge checkable rather than something you have to take on trust.
Token usage and cost appear in Usage, broken down the same way as self-hosted.
Settings → Billing → Model spend lists each charge on one line, even when the month's grant and your purchased balance shared it: tokens for a model call, N search units for a rerank, N pages for scanned pages read.