We worked hard to bring this to production. Would love to have you all test it and give us feedback :)
PS: @aljaz@semisol, thanks for recommending Tinfoil! I mention both of you in the article.
Insane that Fable doesn't write to cache after the initial system prompts. Notice that it wrote 4.9k tokens to cache when the session began and never again. I feel rugged! This planning session cost me 8.9k sats. 😭
Deepseek V4 Pro caches everything and costs much less, proportionally. (see image below)
ps: this is routstrd top that shows cache r/w data
Better defaults - GLM 5.2 and Kimi K2.7
Better routing/switching
Better cache hits
This is also what we think about for Routstr. Our goal is to help the community here save on AI token spends.