Samaya AI unveils sterling, a cost-effective financial research AI model outperforming claude fable 5.1

Samaya AI has introduced Sterling, a financial research agent model that outperforms Claude Fable 5.1 on the FrontierFinance benchmark while operating at 11 times lower cost. Built on NVIDIA’s Nemotron 3.5 Lightning, Sterling is a LoRA post-trained model with 3B active parameters. It scores 58.9% on the benchmark, compared to Fable 3.1’s 55.9%, at a cost of $0.61 per query. Maithra Raghu, in an X post, noted the project began as a two-person effort focused on agentic orchestration efficiency. Sterling’s efficiency stems from a tailored training approach, including a production finance harness adapted for a small model, condensed tool results, and a research budget to limit agent activity. Training used LoRA on Tinker, with performance improving across four stages, reaching 58.9% accuracy. The model itself costs $0.04 per query, with the rest of the cost attributed to tools and helper models. Samaya’s high-effort system achieves 60.6% accuracy at $1.83 per query, demonstrating its potential in complex financial tasks.