Aider has achieved a new state-of-the-art score of 64.0% on its polyglot benchmark by pairing DeepSeek R1 as the Architect model with Claude Sonnet as the Editor.
- The combination outperforms previous results while costing 14 times less than the prior SOTA set by o1.
- Pairing o1 with Sonnet did not yield better results than using o1 alone, unlike earlier thinking models which improved when paired with various editors.
- The benchmark results utilize R1's standard final output rather than its reasoning tokens, as experiments with thinking tokens performed worse.
This configuration demonstrates that specific model pairings can significantly enhance coding performance and cost-efficiency compared to using single reasoning models.