A hard budget on
every LLM key.
LLM spend arrives the way cloud spend did: one key, shared, unbudgeted, and surprising at month end. Issue a virtual key per team with a USD cap and a model allow-list, and read the spend next to your cloud costs.
OpenRouter
Budgets first, then visibility.
A cap that actually stops spend beats a dashboard that reports it afterwards.
Connect a provider
OpenAI, Anthropic, OpenRouter or AWS Bedrock. The credential is verified against the vendor before it is stored, so a revoked or mistyped key fails at connect instead of silently going green and failing on someone's first inference.
Issue keys with caps
One key per team, each with a hard USD budget and the list of models it is allowed to call. Rotation overlaps, so a key can be replaced without breaking a running deploy. Every key is stamped with whoever created it.
Route by complexity
A router classifies each prompt to a cheap or a strong tier on length, max tokens and keyword rules. The report tells you what percentage was served by the cheap one, which is the number that moves the bill.
Read the spend
By provider, model, tier and team, with average latency and failed request counts, sitting beside your cloud spend in Cost Reports. During a brief upstream outage it reports stale figures rather than a confident zero.
What ZopNight does not touch.
Cost governance does not require reading anyone's prompts.
- Content
- ZopNight never processes or stores LLM request or response content. Completions go to the gateway endpoint directly. There is no prompt logging and no content inspection.
- Tenant isolation
- Each deployment is registered under an org-scoped name and each virtual key carries an alias mapping the friendly name onto that org's own deployment. Your team still sends
model="gpt-4o"and reaches only your credential. - Budget enforcement
- Per-key maximum budget, plus an org-wide ceiling across every key. The aggregator reads live spend and fires a warning at 80 percent and an exceeded alert at 100, naming the key. The gateway's own 429 is the hard stop.
- Access control
- Virtual keys and AI models are two separate capabilities in the role matrix, each with view, create, edit and delete, and each scopeable to a single provider. Registering a model connects a credential, so it is admin-only to mutate.
Cap it before the invoice arrives.
Connect one provider, issue one key with a budget, and see where the spend actually goes.