# Anthropic

> ZopNight registers 3 Anthropic models through the AI Gateway: claude-sonnet and claude-opus as strong tiers, claude-haiku as the cheap tier. Each maps to the vendor's `-latest` alias, so a model upgrade lands without reconfiguration. Spend is metered per request against per-team budgets.

Source: https://zop.dev/integrations/ai-gateway/anthropic
Updated: 2026-08-20

---

## Common questions

### Which Claude models does ZopNight register?

Three: claude-sonnet and claude-opus in the strong tier, claude-haiku in the cheap tier.

### Can I pin a specific Claude version?

The defaults map to the vendor's -latest aliases, so an upgrade reaches your traffic with no config change. That is convenient and it is a tradeoff, because the model behind a name can shift under you. Pin explicitly when you need reproducibility.

### Does ZopNight see my Claude prompts?

No. The gateway runs on hosted LiteLLM, which ZopNight configures, meters and budgets, but prompt and completion bodies do not traverse ZopNight infrastructure.

## Claude models aliased to -latest

Three, each mapped to a `-latest` alias rather than a pinned version:

- **Strong**: `claude-sonnet` → `claude-3-5-sonnet-latest`, `claude-opus` → `claude-3-opus-latest`
- **Cheap**: `claude-haiku` → `claude-3-5-haiku-latest`

Aliasing to `-latest` means a vendor upgrade reaches your traffic without a config change. That
is convenient and it is a tradeoff: the model behind a name can change under you, so pin
explicitly if you need reproducibility.

## Your Claude prompts never reach ZopNight

ZopNight never sits in your request path. The gateway runs on hosted LiteLLM; ZopNight
configures it, meters it and enforces the budget, but your prompts and completions do not
traverse our infrastructure. That distinction is why per-request cost attribution is possible
without us reading request bodies.

## Per-team Anthropic keys with USD caps

Each team gets a virtual key with a hard USD budget and its own model allow-list. When a key
hits its cap it stops working. This is a limit, not an alert. Keys rotate and revoke
independently, so a leaked key is contained to one team's budget rather than the org's.

## Routing Claude traffic between strong and cheap

Three postures decide which model serves a request: **Cost-saver**, **Balanced** and
**Quality-first**. Requests are classified into four semantic categories (coding, analysis and
reasoning, extraction and classification, chit-chat) and routed accordingly, so cheap models
take the traffic that does not need an expensive one.

A semantic cache sits in front, with strictness settings from strict (0.98 similarity) to loose
(0.90). On repetitive workloads the cache is usually a larger saving than the routing itself.