Report
Jev sub-calls reportedly cost half as much and run 6× faster than Haiku 5.5 on OOLONG at 1M tokens
A user says they swapped Jev into recursive language models' sub-calls and examined abstention and routing to bigger models.
TLDR
A user says replacing the sub-calls in recursive language models with Jev made the setup half the cost and 6× faster than Haiku 5.5 on OOLONG at 1M tokens. They also mention findings on abstention, routing to bigger models and cases where Jev did not help.
Combined views
398
2 Sources, first seen ago
