58.9% vs 55.9%, $0.61 per Query vs $6.89: Financial Research Models Are Now Competing on Cost of Ownership
Samaya AI's Sterling scores 58.9% on the FrontierFinance benchmark at a cost of $0.61 per question, while Claude Fable 5.1 scores 55.9% with a per-question cost of $6.89. The two models are separated by just 3 percentage points in performance, and Sterling's cost is roughly 1/11 of its rival. The truly noteworthy metric for the industry is the per-unit operating cost in financial research scenarios.
As reported by Asif Razzaq at MarkTechPost, Samaya AI has released Sterling, a post-trained model built specifically for financial research use cases. It is post-trained on top of Nemotron 3.5 Lightning, and the original announcement describes it as a 3B-Active Agent model tailored for financial research. The public performance figures are very straightforward: on the FrontierFinance benchmark, Sterling achieves a 58.9% score at $0.61 per query, while Claude Fable 5.1 scores 55.9% at $6.89 per query. Sterling outperforms the latter by 3 percentage points, and its cost is roughly 1/11 that of Claude Fable 5.1.
Pairing these two sets of numbers together tells a far more meaningful story than simply looking at benchmark leaderboard rankings. Financial research queries often require repeated fact-checking and cross-verification, a workflow that further amplifies per-question costs. The gap between $0.61 and $6.89 quickly translates to a critical budget constraint in scenarios that require batch processing of large numbers of queries. A 3-percentage-point lead on benchmark scores is certainly meaningful, but a full order of magnitude lower in costs could determine far earlier whether a model can be integrated into regular day-to-day workstreams.
New model releases increasingly come with two sets of metrics: a benchmark scorecard, and a cost breakdown. Sterling has laid both of these on the table this time. The original announcement puts the spotlight on the FrontierFinance performance and the 11x cost advantage, which signals that competition in financial AI has begun to shift from "who is smarter" to "who delivers better value for money". This conclusion comes with clear caveats: the public disclosures only provide benchmark scores and per-query costs, and do not cover metrics that investment research teams prioritize far more, such as latency, context window length, tool calling capabilities, and citation accuracy. It is still too early to draw a conclusion on whether the model can replace existing production workflows.
What is certain is that financial research models can no longer be evaluated by performance scores alone. For developers, the $0.61 per-query cost means the model's operating cost is low enough to be factored directly into product design. For regular end users, this means they can run far more queries with the same fixed budget. The next step is to evaluate its real-world performance on earnings report Q&A, data cross-verification and long document processing, rather than fixating solely on a single line of result from the FrontierFinance benchmark.
Original report link: MarkTechPost Coverage
发布时间: 2026-10-30 15:26