Samaya, a startup building AI research agents for finance, said on 28 October 2026 that its new model, Samaya Sterling, scores 58.9% on FrontierFinance at $0.61 per question, against 55.9% at $6.89 for Claude Fable 5.1, according to its blog post and its leaderboard page. FrontierFinance is Samaya’s own benchmark.
Sterling is a post-trained version of NVIDIA’s open-source Nemotron 3.5 Lightning, a 30B mixture-of-experts model with 3B parameters active at a time. By our arithmetic, $6.89 divided by $0.61 is about 11.3 times, close to the “11ร decrease in cost” Samaya states.
Short explainer · Samaya’s 3B-Active Model Beats Fable 5.1 at 11x Lower Cost
A vertical news Short narrated from this piece’s headline and opening line.
Business Pill · The Model and the Harness
An evergreen Business Engineer explainer on models and harnesses. It teaches the general idea only and does not discuss Samaya or this result.
The key insight: Samaya says its post-trained Sterling model, with 3B parameters active, scores 58.9% on its own FrontierFinance benchmark at $0.61 per question, against 55.9% and $6.89 for Claude Fable 5.1; by our arithmetic that is about 11.3 times cheaper, with the costs counted differently and the two run in different harnesses.
Samaya says Sterling “satisfies 58.9% of rubric items on the 220 public questions of FrontierFinance”, which it describes as “3.0 pp above Claude Fable 5.1”. It says training lifted the base model from 42.5% to 58.9%, and that the model’s median number of tool calls per question rose from 3 to 17.
On speed, Samaya gives 282 seconds per question for Sterling against 268 seconds for Fable 5.1, which it calls “a similar latency”. By our arithmetic, Sterling is 14 seconds, or about 5%, slower per question. In a high-effort mode, Samaya says Sterling reaches 60.6%.
Cost per question on Samaya’s FrontierFinance benchmark, with each system’s score, from Samaya’s leaderboard (28 Oct 2026); the 11.3x ratio is our arithmetic, and Sterling’s cost counts tools and helper models while the others count only the agentic model.
How It Was Built
Samaya says it trained the model with LoRA on Tinker, Thinking Machines Lab’s training API, in four stages: two supervised learning stages that learn from “eight diverse teachers”, then offline and online reinforcement learning. It reports scores of 51.4% after synthetic SFT, 54.8% after best-teacher SFT, 57.5% after offline GRPO and 58.9% after online GRPO.
The company says the model and its research harness were trained together: “we train and evaluate in the harness we serve”. It adds that early results with a larger model on the same recipe show more gains, to be shared in a future post.
What the $0.61 Covers
The two cost figures are not measured the same way. Samaya’s leaderboard says Sterling’s reported cost “includes the research model, tools and helper models, with cached agent tokens”, while other systems’ cost calculations count only the agentic modelcost and exclude related costs such as the search API and data-index pricing.
Samaya says that with its trained model “the research costs $0.04 per question”, against $0.56 to $0.95 for frontier closed models doing the same research in the same harness. By our arithmetic, the $0.04 is about 7% of Sterling’s $0.61; the rest, Samaya says, goes to tool calls and the helper models.
How It Fits Our Coverage
Samaya is not the only company this week to publish results from a model it trained for its own product. Canva said its own designmodel cuts edit costs 46% against its previous agent, and Figma said it trained Llama on 250,000 examples.
We covered Claude Fable 5.1’s debut at the top of Artificial Analysis’s Intelligence Index. On Samaya’s leaderboard, Fable 5.1 is the highest-scoring frontier model, at 55.9%.
The Structural Read
As we read Samaya’s post, the result is about a model and a harness trained together: the company says it trains and evaluates in the harness it serves, and that harness changes alone move scores.
As we read the cost figures, most of Sterling’s $0.61 is not the model: Samaya puts the research itself at $0.04 per question, about 7% by our arithmetic, with the rest going to tool calls and helper models.
As we read the leaderboard, the comparison is Samaya’s own system in its own harness against Fable 5.1 in the open-source Finance Agent v2 harness, so it is a product-level comparison rather than a model-for-model test.
Samaya, blog post of 28 October 2026
“post-training even a much smaller model results in a meaningful increase in quality over Fable 5.1 and an 11ร decrease in cost”
Three Implications
FOR SAMAYA The company says early results with a larger model trained on the same recipe show more gains, which it plans to share in a future post.
FOR FRONTIER LABS On Samaya’s leaderboard, Claude Fable 5.1 is the highest-scoring frontier model at 55.9%, and the leaderboard lists its cost at $6.89 per question in the Finance Agent v2 harness.
FOR BUYERS OF AI RESEARCH TOOLS Samaya’s leaderboard says use-case and rubric-category breakdowns for Sterling have not been published, and we have not seen an independent replication.
The framework’s thesis: “The agentic surface — the harness — is becoming where value pools in the entire AI stack”.
As we read it, Samaya training a small model inside its own finance harness is one example of that idea at the vertical level. The framework does not discuss Samaya or FrontierFinance.
What Is Not Established
FrontierFinance is Samaya’s own benchmark, and the two headline systems were not run in the same setup. Samaya’s leaderboard lists Fable 5.1’s 55.9% and $6.89 under “the open-source Finance Agent v2 harness”, while Sterling runs in what the page calls “the proprietary Samaya harness”, which the company says it adapted for the small model.
The leaderboard also says use-case and rubric-category breakdowns for Sterling have not been published. We have not seen an independent replication, and the pages we read do not say whether Sterling’s weights will be released.
This Week In The AI Supercycle
Execution Fault Lines, Concentrated Demand, and Geopolitical Firewalls
This week’s issue: the AI build is not short of demand for its parts. It is short of cheap money, and its end revenue rests on a narrow base of large, often venture-funded buyers.
Samaya says its 3B-active Sterling model scores 58.9% on its FrontierFinance benchmark at $0.61 per question, against 55.9% and $6.89 for Claude Fable 5.1, about 11 times cheaper by our arithmetic. The two systems ran in different harnesses, the costs are counted differently, and the benchmark is Samaya’s own.
A note on sourcing. The model, training, score, cost and latency figures come from Samaya’s own blog post and FrontierFinance leaderboard page, both dated 28 October 2026; FrontierFinance is Samaya’s benchmark and we have not seen an independent test. We did not contact Samaya. Figures labelled our arithmetic are our own calculations from those numbers.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
