Samaya AI has released Samaya Sterling, a finance research agent model that activates just 3B parameters and still scores 58.9% on the company’s FrontierFinance benchmark. Claude Fable 5.1 lands at 55.9% on the same test. Samaya says Sterling answers each question for $0.61, against $6.89 for Fable 5.1, which works out to roughly 11x lower cost. MarkTechPost and nokiapoweruser.com both covered the launch, and Samaya’s own engineering blog, dated October 28, 2026, carries the full numbers.
The headline is easy to repeat and harder to weigh. A small, specialized model beating the strongest frontier system on a leaderboard sounds like a David-and-Goliath story, but the details matter: who built the benchmark, which harness each model ran inside, and where the money actually goes. This analysis walks through the training recipe, the scores, the cost math, and the caveats a finance team should check before moving a research workflow onto a new model.
Sterling is a post-trained version of Nvidia’s Nemotron 3.5 Lightning, a 30B mixture-of-experts model with 3B active parameters. It belongs to the same Nemotron 3.5 generation as the open-weight model we covered in Nemotron 3.5 Transcribe with 40 languages, though the two models do very different jobs. Samaya trained Sterling to plan research, call tools, gather evidence, and write a report.
What Samaya Released and When
Samaya published the Sterling write-up on October 28, 2026, under the title “Samaya Sterling: A post-trained model for state-of-the-art financial research”. MarkTechPost ran its coverage two days later, on October 30, with a headline that put the three selling points up front: 3B active parameters, a win over Claude Fable 5.1 on FrontierFinance, and an 11x cost advantage. HuggingNews framed it the same way, calling it a 3B parameter model that beats Claude on a financial research benchmark.
Samaya describes Sterling as its first post-trained model for financial research. Instead of wrapping a large lab’s model in tooling, Samaya took a small open model and taught it the specific job of an analyst.
Availability is narrow for now. According to the Samaya blog, Sterling runs inside the Samaya platform, and customers need to book a demo or request access. The post does not describe a public API or downloadable weights, so developers should not expect to pull Sterling from a model hub today. The FrontierFinance dataset is a different story. Samaya’s benchmark page says the dataset is open-source and hosted on Hugging Face.
Inside Sterling’s Architecture: A 30B MoE With 3B Active
The architecture choice drives most of the cost story. A mixture-of-experts model stores many parameters but routes each token through only a few expert blocks. Sterling carries 30B parameters in total and uses 3B of them per token. Inference cost and speed track the active count, so a 3B-active model behaves like a small model at serving time while keeping more stored knowledge than a dense 3B network could.
The base model and the training method
Samaya trained Sterling with LoRA on Tinker, the training API from Thinking Machines Lab. LoRA updates a small set of adapter weights instead of the whole network, which keeps the compute bill low and makes repeated experiments affordable. For a company iterating through four training stages, that matters as much as the final score.
The harness: 37 tools and four helpers
The model does not work alone. Samaya’s harness gives Sterling 37 finance-specific tools backed by in-house retrieval models. Around those tools sit four components. A condenser shortens long tool results before they reach the model. A compactor summarizes the research trace as it grows. A research budget caps how long the agent keeps digging. A report writer turns the gathered evidence into the final answer. MarkTechPost summarized the core idea as co-training the model with its harness, and the blog says trimming the tool list and compressing long results lets a 3B-active model sustain research that would otherwise overflow its context window.
If you build agents yourself, the tool-count point is worth a pause. Handing a small model dozens of loosely described tools usually hurts accuracy. Our function calling tutorial across 4 LLMs walks through the schema discipline that makes tool use reliable, and Sterling’s results show how far that discipline can carry a small model once the training data matches the tool set.
The Four-Stage Training Recipe
Samaya post-trained Sterling in four stages: two rounds of supervised fine-tuning, then offline and online GRPO. GRPO is a reinforcement learning method that scores groups of sampled answers against each other, and it has become a common tool for teaching open models to reason. Samaya’s version rewards the model for the share of essential rubric items its report satisfies.
Stage one used 968 synthetic analyst questions. A teacher model from the Kimi family, Kimi K3, answered each question twice, and an LLM judge picked the better answer. No rubrics entered this stage. Its job was to teach the mechanics of calling the 37 tools and collecting evidence. For readers tracking the newer Kimi release, our Kimi K3.1 Pro comparison covers how it stacks up against other frontier systems.
Stage two switched to expert-written material. Samaya ran 1,819 expert-written questions against eight different research strategies and kept the best-graded answer for each. That selection lifted average rubric satisfaction from 66% to 80%. The team then trained for five epochs on 1,157 questions, and the third epoch scored best on validation. Stage three, offline GRPO, reused all eight teacher traces and ran 132 steps of 8 questions with 8 answers each. Stage four, online GRPO, let Sterling answer each training question 8 times itself, over 26 steps of 16 questions. Samaya focused that stage on questions where teacher scores varied and on budget-constrained examples, and it applied truncated importance sampling to correct for the mixture-of-experts routing.
| Training stage | Method | FrontierFinance score |
|---|---|---|
| Base model (Nemotron 3.5 Lightning) | No post-training | 42.5% |
| Stage 1 | Synthetic SFT, 968 questions | 51.4% |
| Stage 2 | Best-of-eight SFT, 1,157 training questions | 54.8% |
| Stage 3 | Offline GRPO, 132 steps | 57.5% |
| Stage 4 | Online GRPO, 26 steps | 58.9% |
Source: Samaya AI engineering blog, October 28, 2026. Scores come from the 220 public FrontierFinance questions, which Samaya says stayed strictly held out from training. A separate set of 100 internal questions served as validation throughout.
FrontierFinance Scores: 58.9% Against 55.9%
The core result is a 3.0 percentage-point lead. Sterling satisfies 58.9% of rubric items on the 220 public FrontierFinance questions, while Claude Fable 5.1 satisfies 55.9%. Samaya calls Fable 5.1 the strongest frontier model on the leaderboard, so the comparison targets the best available rival, not a weak baseline.
The training curve tells a more interesting story than the final gap. Post-training added 16.4 points to the base model, from 42.5% to 58.9%. Every stage contributed, and the reinforcement learning stages alone added about 4 points on top of the supervised result. Behavior changed too. The median number of tool calls per question climbed from 3 to 17 across training, which suggests the model learned to research instead of answering from memory.
The gains hold up on hard questions. On the hardest third of the benchmark, Sterling moved from 31.8% for the base model to 45.2%, a 13.4-point jump. Samaya reports consistent improvement across all difficulty levels, which argues against a result driven only by easy questions.
The 11x Cost Gap, Priced Out
Cost per question is the number that turns heads. Samaya lists Sterling at $0.61 and Claude Fable 5.1 at $6.89, a ratio of about 11.3 to 1. Scale that to the full public benchmark of 220 questions and the arithmetic gives roughly $134 for Sterling against roughly $1,516 for Fable 5.1. An analyst desk running a few thousand research queries a month would see the gap show up in the budget line quickly.
One detail deserves attention. Samaya breaks out the model’s own cost at $0.04 per question, separate from tools and helper components. That implies most of the $0.61 sits in the harness, in retrieval and supporting calls, and not in the 3B-active model itself. For buyers, the lesson is that shrinking the model only lowers one part of an agent’s bill. The rest depends on how many tool calls the agent makes, and Sterling makes 17 at the median.
Token pricing for frontier models keeps falling at the low end, so the 11x figure will age. Our look at Haiku 5.5, GPT-6.1 Terra, and DeepSeek API pricing shows how cheap capable general models have become, and our Claude Opus, Sonnet, and Haiku 5.5 price comparison lays out the Anthropic ladder. A fair test would pit Sterling against a cheaper frontier tier inside the same harness, and Samaya’s published table does not include that run.
Latency and the High-Effort Mode
Speed does not favor Sterling, though the gap is small. The standard configuration takes 282 seconds per question, while Claude Fable 5.1 takes 268 seconds. Sterling runs about 14 seconds, or 5%, slower. Neither number suits a chat window. Both describe long-running research jobs where a user submits a question and returns minutes later.
Samaya also reports a high-effort setting. It scores 60.6% at $1.83 per question and 372 seconds. That is 4.7 points above Fable 5.1 at about a quarter of the cost, but it adds roughly 104 seconds of waiting. Teams can pick the tradeoff per task. A morning screen across fifty names might use the cheap setting, while a deep dive before an investment committee meeting could justify the longer run.
Competitive Comparison: Sterling vs Fable 5.1 vs Muse Spark Open
The FrontierFinance leaderboard on Samaya’s benchmark page puts Sterling first. Claude Fable 5.1 running in a Finance Agent v2 harness sits third at 55.9%. Muse Spark Open, also in the Finance Agent v2 harness, scores 53.7% at $1.38 per question and 234 seconds, making it the fastest of the top entries. The table below combines the leaderboard with the figures from Samaya’s blog.
| System | FrontierFinance score | Cost per question | Latency |
|---|---|---|---|
| Samaya Sterling (high effort) | 60.6% | $1.83 | 372 sec |
| Samaya Sterling (standard) | 58.9% | $0.61 | 282 sec |
| Claude Fable 5.1 (Finance Agent v2) | 55.9% | $6.89 | 268 sec |
| Muse Spark Open (Finance Agent v2) | 53.7% | $1.38 | 234 sec |
| Nemotron 3.5 Lightning base model | 42.5% | n/a | n/a |
Samaya argues that Sterling Pareto-dominates every high-quality system on the leaderboard when you plot quality against cost. In plain terms, no listed system delivers a better score at a lower price. The table supports that reading for the entries shown. Muse Spark Open costs more than twice as much as Sterling and scores 5.2 points lower, and Fable 5.1 costs more than eleven times as much for 3.0 points less.
The comparison also exposes a harness effect. The big-model entries ran inside the Finance Agent v2 harness, while Sterling ran inside the harness it was co-trained with. That is the point of Samaya’s approach, but it means the leaderboard measures model-plus-harness pairs, not raw model intelligence. A different harness could move Fable 5.1’s number in either direction.
How FrontierFinance Works, and Where It Falls Short
FrontierFinance contains 220 queries graded against 11,543 rubric items, an average of 52.5 items per query. Experts wrote the rubrics. One example on the benchmark page, a company-research question about Marqeta, carries 44 rubric items, 13 of them marked must-have, covering business model, revenue mechanics, financial metrics, and margins. A model’s score is the share of rubric items its answer satisfies.
| Workflow | Share of benchmark | What it tests |
|---|---|---|
| Discover | 8% | Screening and discovery |
| Research | 15% | Company research |
| Evaluate | 49% | Sector and macro (17%), financial data extraction (32%) |
| Monitor | 28% | Coverage and catalysts (12%), earnings and events (16%) |
The mix leans toward evaluation work. Nearly half the questions ask for sector views or data extraction, with data extraction alone at 32%. A model that excels at pulling numbers from filings will look strong here, and one built for long-form qualitative judgment may look weaker than it is. Buyers should check that the mix matches their own workload.
The larger limitation is provenance. Samaya built FrontierFinance, and Samaya trained Sterling. The company published the dataset openly, which lets outsiders audit the questions, and it kept the 220 public questions out of training. Even so, a vendor-run benchmark deserves independent replication before anyone treats the ranking as settled. Rubric grading also depends on an LLM judge, and judges carry their own biases.
Why Co-Training the Model With Its Harness Matters
Most agent products today treat the model and the scaffolding as separate layers. A team picks the best available model, writes prompts and tool definitions, and tunes both by hand. Sterling inverts that order. The tools, the condenser, the compactor, and the budget all existed during training, so the model learned to work with their quirks. Its jump from 3 to 17 median tool calls shows it learned to lean on them.
This approach has a price. A model trained against one harness may transfer poorly to another, and Samaya has not published results showing Sterling inside a rival’s scaffolding. It also means the moat sits in proprietary assets: expert-written questions, rubrics, in-house retrieval models, and a tool set refined over years. A competitor cannot copy the recipe from the blog post alone.
Historical Context: From General Chatbots to Specialist Agents
The first wave of enterprise AI in finance leaned on general chat models with retrieval bolted on. Vendors differentiated on data access and interface, since every product sat on the same handful of frontier models. Rising API bills then pushed buyers to ask whether a smaller model could do the same job, and the open-weight ecosystem supplied the raw material.
Reinforcement learning against verifiable rewards changed the economics of specialization. Rubric scoring gives finance teams something like a verifiable reward, because experts can decide ahead of time which facts an answer must contain. Sterling applies that idea to a research agent and shows it can beat a general frontier model on a narrow domain. Other sectors with expert-written rubrics, such as law, medicine, and tax, offer the same opening.
Market Impact for Finance Teams and Model Vendors
For finance teams, the near-term effect is pricing pressure on research tooling. If a 3B-active model matches or beats a frontier model on analyst tasks, vendors who resell frontier tokens at a markup lose cover for their margins. Samaya’s own position improves too, since a model it trained itself gives it direct control over serving costs.
For frontier labs, the result is a reminder that general leadership does not guarantee leadership in every vertical. Anthropic, OpenAI, and Google still set the ceiling on broad reasoning, and our AI and machine learning coverage tracks those releases closely. But a vertical specialist with proprietary data and a tuned harness can win a specific benchmark at a fraction of the cost. Expect more labs to pitch fine-tuning and agent-harness services to defend premium accounts.
Nvidia also gains. Sterling ships on a Nemotron base, which adds evidence that the Nemotron 3.5 line can serve as a foundation for serious domain work. Each such result makes the open Nemotron family a stronger choice for enterprises that want to own their weights and tune them in-house.
What to Check Before Trusting the Numbers
- Run your own questions. Build a small rubric set from your team’s real research tasks and score both systems blind.
- Compare harnesses fairly. Ask what scaffolding each model ran in, and test the frontier model inside the best harness you can build.
- Count total cost. Include tool calls, retrieval, and retries. Samaya’s model-only figure of $0.04 per question shows how much sits outside the model.
- Measure latency against your workflow. A 282-second answer suits overnight batches but not live client calls.
- Look at the hard tail. Sterling’s 45.2% on the hardest third means more than half of those rubric items still go unmet.
- Ask about access. Sterling requires a demo request, so confirm data residency, audit logs, and contract terms early.
Five Predictions
These are analyst judgments, not reported facts, and any of them could miss.
- Independent replication arrives within months. The open FrontierFinance dataset invites outside teams to test Sterling rivals, and at least one lab or startup will likely publish its own scores.
- Frontier labs answer with vertical harnesses. Expect finance-specific agent products from the large labs that bundle tools, retrieval, and tuned prompts to close the gap on cost and score.
- More vendors post-train small MoE models for their own domains. Legal, tax, and healthcare research vendors have rubric data too and will copy the pattern.
- Benchmarks start reporting total cost and harness details by default. Buyers have learned that a score without a cost and a scaffold tells little.
- Samaya widens access. A demo-only model limits adoption, so an API tier or partner channel is a natural next step, though the company has announced nothing of the kind.
The Takeaway
Samaya Sterling shows that a 3B-active model, trained against the exact tools and rubrics it will face, can edge out the strongest frontier system on a finance research benchmark while costing about one-eleventh as much per question. The 3.0-point lead is modest, the benchmark belongs to the vendor, and the harness does a lot of the work. Still, the training curve, the open dataset, and the cost breakdown give buyers enough detail to test the claim themselves. That is more than most launch announcements offer.
FAQ
What is Samaya Sterling?
Samaya Sterling is a finance research agent model from Samaya AI. It is a post-trained version of Nvidia’s Nemotron 3.5 Lightning, a 30B mixture-of-experts model with 3B active parameters, trained to plan research, call 37 finance tools, and write reports.
How does Sterling compare with Claude Fable 5.1 on FrontierFinance?
Sterling scores 58.9% on the 220 public questions, against 55.9% for Claude Fable 5.1. Its high-effort setting reaches 60.6%. Both figures come from Samaya’s own benchmark and blog.
Where does the 11x lower cost figure come from?
Samaya lists $0.61 per question for Sterling and $6.89 for Claude Fable 5.1. Dividing the two gives about 11.3. The model’s own share of Sterling’s cost is $0.04 per question, with the rest tied to tools and helper components.
Can I download Sterling or call it through an API?
Not based on the announcement. Samaya says access runs through its platform, with a demo or access request required. The FrontierFinance dataset is open-source on Hugging Face, but the post does not describe public Sterling weights.
How was Sterling trained?
In four stages using LoRA on Tinker: synthetic supervised fine-tuning on 968 questions, best-of-eight supervised fine-tuning on expert-written questions, offline GRPO for 132 steps, and online GRPO for 26 steps. The score rose from 42.5% for the base model to 58.9%.
Is the benchmark independent?
No. Samaya created FrontierFinance and also built Sterling. The company published the dataset and says the 220 public questions stayed out of training, but outside replication would strengthen the result.
Is Sterling faster than Claude Fable 5.1?
No. Sterling takes 282 seconds per question in standard mode and 372 seconds in high-effort mode. Claude Fable 5.1 takes 268 seconds.