1H Caching for Custom Model Provider | Friends of the Crustacean 🦞🤝 | Page 1

mental ridge Apr 22, 2026, 2:57 AM

#

Is 1-hour caching available for the Anthropic message API model?

barren hazelBOT Apr 22, 2026, 2:57 AM

#

bright ruinBOT Apr 22, 2026, 2:57 AM

#

To help others find answers, you can mark your question as solved via Right click solution message -> Apps -> ✅ Mark Solution

barren hazelBOT Apr 22, 2026, 2:57 AM

#

ruby quiverBOT Apr 22, 2026, 3:00 AM

#

Yes, with an important caveat.

For Anthropic Messages, cacheRetention: "long" becomes a real 1 hour TTL only when the request goes to a direct Anthropic host, api.anthropic.com, or Anthropic Vertex.

If you are using a custom/proxy host for an anthropic-messages model, OpenClaw will not upgrade it to 1 hour. It stays on the normal ephemeral cache behavior instead.

Docs:
https://docs.openclaw.ai/reference/prompt-caching

Relevant bit from the docs:

cacheRetention: "short" = default 5 minute ephemeral cache
cacheRetention: "long" = 1 hour, only on direct Anthropic or Vertex routes

If you want, I can help check whether your custom provider route qualifies.

mental ridge Apr 22, 2026, 4:11 AM

#

ruby quiver Yes, with an important caveat. For Anthropic Messages, `cacheRetention: "long"`...

Is there a roadmap for implementing a one-hour cache for custom providers?

ruby quiverBOT Apr 22, 2026, 4:13 AM

#

I don’t see a public roadmap item for 1 hour TTL on arbitrary custom providers right now.

What I do see:

General prompt caching for custom anthropic-messages providers was requested and later completed: https://github.com/openclaw/openclaw/issues/37325
An older related issue was closed as not planned: https://github.com/openclaw/openclaw/issues/19279

But for the specific 1 hour part, the current docs and code still intentionally gate that to known endpoints only:

direct Anthropic, api.anthropic.com
Anthropic Vertex

Docs: https://docs.openclaw.ai/reference/prompt-caching

So current state looks like:

custom anthropic-messages providers can get prompt caching support
cacheRetention: "long" does not currently become a real 1 hour TTL on arbitrary custom hosts

#

If your provider really supports 1 hour Anthropic-style cache TTL, the best next step is probably a GitHub feature request with the provider docs linked as evidence.

mental ridge Apr 22, 2026, 4:51 AM

#

ruby quiver I don’t see a public roadmap item for **1 hour TTL on arbitrary custom providers...

@ruby quiver Is this pull request live now? https://github.com/openclaw/openclaw/pull/59049

ruby quiverBOT Apr 22, 2026, 4:52 AM

#

Yes, that PR is merged and live on main, and it shipped in v2026.4.2 and later.

So if you’re on any recent 2026.4.x build, you should have it.

Small caveat though: PR #59049 makes custom anthropic-messages providers honor explicit cacheRetention values. It does not mean arbitrary custom hosts now get an automatic real 1 hour Anthropic TTL. Current docs/code still only guarantee that for direct Anthropic and Anthropic Vertex.

If you want, paste your openclaw --version and I can sanity check whether your install definitely includes it.

mental ridge Apr 22, 2026, 4:54 AM

#

ruby quiver Yes, that PR is merged and live on `main`, and it shipped in **v2026.4.2** and l...

Current version 2026.4.20

ruby quiverBOT Apr 22, 2026, 4:55 AM

#

Yep, 2026.4.20 definitely includes that PR. You have the fix.

So on your version:

custom anthropic-messages providers should honor explicit cacheRetention
but a true Anthropic-style 1h TTL is still only guaranteed on direct Anthropic / Vertex, not arbitrary custom hosts

If you want, paste the relevant models.providers block and I can sanity check whether your config is set up correctly.

#1H Caching for Custom Model Provider