On October 28, 2026, Samaya AI released Sterling, a model post-trained specifically for agentic financial research, together with FrontierFinance, the benchmark it built to test it. The headline number: Sterling scored 58.9 percent on FrontierFinance, ahead of Claude Fable 5.1 at 55.9 percent, while costing roughly one-eleventh as much per question. For businesses that run finance workflows on AI, the more interesting story is not the score itself. It is what it says about where specialist models, trained narrow and priced low, fit next to frontier generalists in real work.
What Samaya announced
Samaya's research post, published October 28, 2026, describes Sterling as its first post-trained model for "state-of-the-art financial research." Sterling is the Samaya System running the company's post-trained 30-billion-parameter model, with the company reporting its benchmark scores as the mean of two runs. Reporting by MarkTechPost on October 30, 2026 adds architecture detail: the model is built on NVIDIA's Nemotron 3.5 and routes through about 3 billion active parameters, which is part of why the per-question cost can fall so far below frontier rates.
The announcement arrived alongside the benchmark itself. FrontierFinance, also introduced by Samaya on October 28, is a 220-question public evaluation of end-to-end financial research, where quality is measured as the share of rubric items an answer satisfies. The rubric framing matters: the benchmark grades finished research work, not single-turn chat, and the comparison systems' scores are drawn from the public FrontierFinance leaderboard.
The results, in a table
| System | FrontierFinance quality | Cost per question |
|---|---|---|
| Samaya Sterling | 58.9 percent | 0.61 dollars (open model serving price plus tools and the models that write and condense, per Samaya) |
| Claude Fable 5.1 | 55.9 percent | Roughly eleven times higher, per Samaya's published comparison |
| Other systems | Public FrontierFinance leaderboard | Leaderboard figures |
Three caveats belong next to that table. The benchmark is Samaya's own, the scores are vendor-reported, and rubric-based grading reflects what one company decided finished financial research looks like. That does not make the result meaningless; it makes it a claim to test, not a fact to adopt.
How a small model beats a frontier model
The technical story is post-training. Sterling was trained after its base model, with reinforcement learning aimed at the specific behaviors financial research demands: working through long chains of tools, retrieving documents, reconciling figures and synthesizing an answer that satisfies a rubric rather than amusing a reader. Frontier models are optimized for breadth across every domain and task type. A specialist concedes generality and buys depth.
The fit is strongest where the work is repetitive and gradable: reconciliation research, vendor and counterparty diligence, drafting board-deck support, triaging anomalies in accounts receivable and payable, and summarizing filings against a checklist. These are exactly the workflows where a rubric exists or can be written, which is what makes the economics interesting: at 0.61 dollars per question, a hundred exploratory analyses cost less than sixty dollars, a level where testing a hypothesis stops being a budget decision. For a small business, a useful starting rule: begin with the workflow your team repeats every week and still grades by hand, because that is where automation pays for itself fastest and errors are cheapest to catch.
What this means for finance workflows in small businesses
Our read, and what we would tell a client: the frontier-vs-specialist question now has a real answer pattern, and it is routing, not replacement. Keep the frontier model for judgment calls, ambiguous briefs and cross-domain work. Send the repeatable, rubric-shaped research to the narrow model and audit its outputs periodically against human review. The cost asymmetry makes tiering a strategy.
That said, a vendor benchmark is a starting gun, not a finish line. Before committing a workflow to any specialist model, ask four questions: who wrote the rubric and does it match your definition of done? Are the scores independently reproducible? How current is the underlying financial data? And what tool access does the system assume, since agentic finance work lives and dies by retrieval? These questions echo the trust gap we documented in October 2026 research from OneStream: enterprises report expanding AI in finance most where they can verify outputs and control data.
A four-week pilot for a small finance team
The cheapest way to act on this news costs almost nothing:
- Pick one rubric-shaped workflow. Vendor diligence, customer credit checks or AR anomaly triage are ideal: repeatable, checklist-like and low-risk to assess.
- Write the rubric before you test anything. Define the checks a good analyst would run, in the order they would run them. Ten items is plenty.
- Run the same thirty to forty recent, real tasks through both a frontier model and the specialist candidate, blind. Grade against your rubric, not vibes.
- Compare two numbers: quality per rubric, and cost per one hundred questions. If the specialist matches quality at a tenth of the cost, route the workflow there and re-run the pilot quarterly.
The discipline that makes this work is the rubric. A benchmark you wrote yourself, even a rough one, beats a vendor benchmark for decisions, because it encodes your definition of done instead of theirs.
Sterling's result will not be the last of its kind. Expect the next year of finance AI to look like this: small, post-trained specialists beating or matching frontier models on defined tasks at a fraction of the cost, and the winning deployment pattern being disciplined routing between them. If you want help mapping which of your finance workflows are rubric-shaped and worth routing, that is a conversation we enjoy. Start at The Vivid Company.
