Anthropic

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

A -latest alias follows new snapshots of the same model version without a config change, but it never moves to a newer Claude generation. The model behind a name can still change between snapshots, so pin explicitly if you need reproducibility.

Where your Claude prompts go

Requests pass through the ZopNight-hosted LiteLLM gateway. ZopNight configures it, meters it and enforces the budget, and the complexity router reads each prompt to choose the cheap or strong tier. Cost attribution works from per-request usage data: key, model, tokens and cost.

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.