Your own Anthropic key now gets prompt caching

Chats on your own Anthropic key were never read from the cache. ZeroLog now uses the native Messages API, so repeated input is billed at the cached price.

Nutzerwirkung: Noch nicht bewertet

If you store your own Anthropic key under Providers and keys, ZeroLog now talks to Anthropic through its native Messages API. Before, every request went through Anthropic's OpenAI-compatible layer, and that layer does not support prompt caching. Each round of a long agent run paid the full input price again, even though the system prompt, the tools and the conversation so far had not changed.

The path of one chat request with your own Anthropic key, before and after the change.

Text version

Before: ZeroLog sent the request to the compatibility layer, which ignores cache marks. Of 10 large requests on own keys in the last 14 days, none was read from the cache. After: ZeroLog sends the request to the Messages API with up to four cache marks, on the system prompt, the last tool, the last message and the last answer before your newest question. Pool models reached 56 of 128 in the same period.

The usage figures now count cached input tokens for every provider that reports them, under each provider's own field name: OpenAI and xAI, Moonshot, DeepSeek and Anthropic. Your own key is still never billed by ZeroLog, so nothing changes on your balance. The saving shows up on your Anthropic invoice.

The other providers you can connect cache automatically once the start of the prompt stays the same. OpenAI, xAI, Together, Mistral, Moonshot and Cerebras also take a cache key, and ZeroLog now sends one per chat. Cached input from DeepSeek and Moonshot was counted as zero before and is read correctly now.

We have not yet seen a request on a real own key after this change. The new code is tested against a replica of the Messages API that caches like the real one, so the first real numbers will appear in the usage list of the next chats on an own Anthropic key.