Samaya AI says its Sterling finance research model scored higher than Claude Fable 5.1 on the FrontierFinance benchmark, at roughly one-ninth the cost per question. The result is a promising example of a smaller, purpose-built model competing through its research tools and workflow, though the benchmark figures are self-reported.
Watch Desk analysis
What happened
Sterling is a post-trained version of NVIDIA’s Nemotron 3.5 Lightning, a mixture-of-experts model with 30 billion total parameters and 3 billion active parameters. It plans research, calls financial tools and gathers evidence before passing its findings to a separate report writer, MarkTechPost reports.
On FrontierFinance’s 220 public questions, Samaya says Sterling scored 58.9%, compared with 55.9% for Claude Fable 5.1. Samaya puts the average cost at $0.61 per question for Sterling and $6.89 for Claude. Both systems took a similar time in the reported comparison: 282 seconds for Sterling and 268 for Claude. The benchmark uses expert-written rubrics and a three-model judging panel, according to the report.
Samaya says the model itself accounts for just $0.04 of Sterling’s per-question cost; tools and helper models make up the rest. In a high-effort mode, Sterling scored 60.6% at $1.83 per question, taking 372 seconds. The reported results come from Samaya’s own benchmark and have not been independently established here.
Why it matters
The interesting proposition is not simply that a smaller model can beat a larger one. Samaya’s account suggests the agent’s tools, context management and research process matter too. Its training combined supervised fine-tuning with two stages of reinforcement learning, using synthetic and expert-written questions.
For teams building research agents, that is a useful reminder: a model’s price tag is only part of the bill, and a carefully designed harness may change the balance between cost and quality. But one company’s benchmark is a starting point for comparison, not a universal verdict on financial research tasks.
Our read
Sterling’s reported cost-quality result is striking, and the similar completion times make the comparison more useful than a cheap-but-glacial win. The sensible next step is to test the model on independent questions and real workflows, where tool use, factual accuracy and the cost of a wrong answer all count. Finance is not a field where “close enough” has much charm.
What to watch
- Whether independent evaluations reproduce Sterling’s cost and accuracy results.
- How it performs on financial questions outside the FrontierFinance benchmark.
- Whether Samaya publishes more detail on Sterling’s serving setup and real-world use.
Discussion spark: Would you trust a lower-cost specialist agent for financial research if it beat a leading model on one benchmark, or would you wait for independent testing?
Sources and evidence
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.