AiderFrom the builder
R1+Sonnet set SOTA on aider’s polyglot benchmark
Aider describes its architect/editor setup, in which R1 plans how to solve a coding task and Claude 3.5 Sonnet turns that plan into concrete file edits. The pairing scored a record 64.0% on the aider polyglot benchmark at roughly one fourteenth the cost of the earlier o1 record ($13.29 versus $186.5). Aider notes the run used only R1's final answers rather than its thinking tokens, and that pairing o1 with Sonnet did not beat o1 alone.
They achieve this at 14X less cost compared to the previous o1 SOTA result.
Saved . A shortened excerpt from the official blog; the original publisher retains authorship.