Samaya AI releases Sterling: a financial agent with 3B active parameters beats Claude Fable 5.1 on price and quality

On October 30, 2026, Samaya AI released Sterling — a model for agentic financial research (agentic research: the model plans its own search, calls tools, and gathers evidence). Sterling is built on NVIDIA Nemotron 3.5 Lightning — a mixture-of-experts (an architecture where only part of the parameters is activated per request) with 30B total and 3B active parameters. Fine-tuning was done using LoRA (low-rank adaptation: adding small trainable matrices to a frozen model). According to Samaya, on its FrontierFinance benchmark Sterling scores 58.9%, while Claude Fable 5.1 scores 55.9%. Cost per question: $0.61 versus $6.89.

The difference from previous approaches lies in coupling the model with its harness (a set of prompts, tool descriptions, and auxiliary calls). Samaya adapted the production harness for a small model: it removed detailed instructions from the system prompt and tool descriptions, merged tools to shorten the prompt, and added two auxiliary modules. As a result, a model with 3B active parameters outperforms a large proprietary model. According to Maitra Raghu in the announcement on X, the project began as a two-person effort on the efficiency of agentic orchestration.

In practice, this is a recipe for developers: jointly fine-tuning a small model and its harness for a narrow subject domain can yield gains in both quality and cost. Sterling does not just answer questions — it plans research, calls financial tools, gathers evidence, and passes findings to a separate module that writes the report. For teams building financial agents, this is an example of how to save on inference without losing accuracy on a specialized benchmark.

It remains unknown how Sterling will behave on other financial tasks beyond FrontierFinance and how reproducible the results are by independent testers. There is also no data on the license, access terms, and full development cost.