Samaya AI says its Sterling finance research model scored 58.9% on the company’s FrontierFinance benchmark, ahead of Claude Fable 5.1’s 55.9%. Samaya reports a cost of $0.61 per question for Sterling, compared with $6.89 for Claude—about 11 times less.
Those figures are Samaya’s own benchmark results, not an independent comparison. FrontierFinance tests research tasks across the investment process, including questions that require gathering financial data and supporting a written answer with citations. Samaya says the benchmark questions were held out from training, but the company created the test and ran the evaluation.
How Sterling was trained
Sterling is based on NVIDIA’s Nemotron 3.5 Lightning, a 30-billion-parameter model that uses 3 billion parameters at a time. Samaya says the base model started at 42.5% accuracy on FrontierFinance before the company’s four-stage training process raised its score to 58.9%.
The process combined two rounds of supervised training with two rounds of reinforcement learning, a method that uses scores to guide later attempts. The model learned to use finance research tools, choose among different research approaches, gather evidence and produce cited reports. Samaya says it trained on examples from eight different teaching models and on questions written and graded by finance professionals.
The research setup matters to the result: Sterling works with a finance-specific collection of tools, and helper models shorten long tool outputs and research histories. Samaya says this setup helps the smaller active model carry out research over many tool calls. The published scores therefore describe the model working within that system, not a standalone model answering without tools.
What the comparison does—and doesn’t—show
The result suggests that post-training and a tailored research setup can improve a smaller model’s performance on Samaya’s particular test. It does not establish that Sterling is more accurate than Claude for finance research in general: the supplied results cover one benchmark, and no independent evaluation is documented.
Samaya says it used a separate set of 100 internal questions to choose training checkpoints. The company also reports that a larger model trained with the same approach has shown early gains, but it has not provided those results yet. The supplied announcement does not specify how people can access Sterling.
Why it matters
For finance teams exploring AI research tools, Samaya’s results offer a reason to examine smaller, tool-equipped models as a potentially lower-cost option. But the scores come from Samaya’s own benchmark, so teams still need independent evidence to know whether Sterling performs as well on their research tasks.
