A small, specialised model has beaten a frontier model on a hard professional task, according to the company that built it. On October 28, 2026, Samaya published Samaya Sterling: A post-trained model for state-of-the-art financial research. It reports that a 30B mixture-of-experts model with only 3B active parameters, trained together with its research harness, scores 58.9% on the FrontierFinance benchmark, against 55.9% for Claude Fable 5.1, at $0.61 per question versus $6.89.
This is the same pattern we have tracked all month: narrow tuning of a small open model beating a general frontier model on one measured job. See Shopify's Sidekick post-trained on Nemotron 3.5 Lightning and Exowatt's fine-tuned Qwen3 beating Claude Opus on alarm triage. Samaya's write-up is unusually detailed about how, so it is worth reading closely, and just as closely for what it does not prove.
TL;DR: Sterling's numbers
| Question | Answer from Samaya's post |
|---|---|
| Base model | NVIDIA Nemotron 3.5 Lightning, open, 30B MoE, 3B active |
| Score on FrontierFinance (220 public questions) | 58.9% (base model: 42.5%) |
| Claude Fable 5.1 | 55.9% at $6.89 per question |
| Sterling cost per question | $0.61 (trained model about $0.04, rest is tools and helper models) |
| Latency | 282 s vs 268 s for Fable 5.1 |
| High-effort mode | 60.6% at $1.83 and 372 s |
| Training | Four stages: two supervised, two reinforcement learning; LoRA on Tinker |
| Teachers | Eight research strategies run through Kimi K3 |
| Independent check | None yet |
What is FrontierFinance, and who made it?
FrontierFinance is a benchmark Samaya introduced in July 2026, described in its post and an arXiv paper (arXiv:2608.11683) as a challenging test of AI agents on workflows across the investment process, from idea screening to macro research. Answers are graded against expert-written rubrics, and the score is the share of rubric items satisfied.
The key fact for any reader: Samaya built the benchmark and the model. That does not make the result wrong, but it means the comparison is chosen by the party that gains from it. Samaya says that FrontierFinance is "strictly held out" from training, that no benchmark question was used for training, and that every checkpoint was chosen on a separate pool of 100 internal validation questions. Those are good practices, and they are the company's account, not an audit.
An example task from the post: "What are the interest coverage ratios for AT&T and Verizon for each quarter from 2019 to 2023? How have their debt service capabilities changed?" Answering takes about 40 company-quarters of data, one consistent formula, and a note on quarters where the ratio misleads.
How was Sterling trained?
Fine-tuning adapters fitting around a small base model, the LoRA approach behind Samaya Sterling
Samaya trained the model with LoRA on Tinker, Thinking Machines Lab's training API, in four stages:
- Supervised fine-tuning on synthetic research traces. 968 synthetic analyst questions; the teacher, Kimi K3, answered each twice with Samaya's tools and an LLM judge chose the better answer. This teaches the base model to call 37 tools with valid arguments and to finish by calling the report writer.
- Supervised fine-tuning on the best of eight strategies. The teacher tried eight research strategies on the same 1,819 expert-written questions. The best-graded answer per question became the training example. Samaya reports that the best of eight satisfied 80% of essential rubric items on average, against about 66% for the single best strategy. After filtering, 1,157 questions were used for five epochs; epoch 3 scored best.
- Offline GRPO on all eight answers. Group relative policy optimization scores each answer against the others for the same question, so good and bad strategies both teach the model. Because the traces already exist, each step is cheap, as in DPO.
- Online GRPO on its own attempts. The model answers each question eight times and a grader scores each against all rubric items. Samaya used truncated importance sampling to correct the mismatch between sampling and training models, a known issue for mixture-of-experts models.
The score climbs at every stage: base 42.5%, synthetic SFT 51.4%, best-teacher SFT 54.8%, offline GRPO 57.5%, online GRPO 58.9%. Imitation adds 12.3 points, reinforcement learning another 4.1.
If GRPO is unfamiliar, our explainer on loop engineering for coding agents covers the surrounding agent-loop vocabulary, and the same recipe family appears in NVIDIA's Nemotron fine-tuning recipe for olympiad-level problems.
Why the harness matters as much as the model
The most transferable line in the post: "the model and harness being co-adapted through the post-training process." The harness gives the model finance tools (the production harness has dozens), and Samaya made it small-model friendly:
- Detailed guidance was removed from the system prompt and tool descriptions, because the model can learn those nuances in training.
- Tools were merged to shorten the prompt.
- A condenser shortens long tool results and a compactor summarises the trace as it grows.
- A research budget caps how long the agent researches, after which a separate report writer answers from what was found.
Samaya notes that small harness changes, such as what the report writer sees, move scores either way, so it trains and evaluates in the harness it serves. For a builder, the lesson is that you cannot compare "model A versus model B" in isolation; the scaffold around the model is part of the system. The median number of tool calls per question rose from 3 for the base model to 17 for Sterling, so a lot of the gain is the model learning to research longer and more systematically.
Reading the cost claim carefully
The headline of 11 times lower cost compares two full systems, not two models. Sterling costs $0.61 per question; Fable 5.1 costs $6.89. Samaya breaks it down: the trained model itself is about $0.04, frontier closed models doing the same research in the same harness cost $0.56 to $0.95, and the rest of the $0.61 goes to tool calls and helper models.
So there are two different comparisons in the post. In the same harness, the model cost gap is roughly $0.04 against $0.56 to $0.95. Against Fable 5.1 as listed on the leaderboard, the full-system gap is $0.61 against $6.89, which presumably reflects a different way of running Fable. The post does not fully reconcile the two in the text we read, so treat the 11 times figure as a leaderboard-level claim and the in-harness numbers as the cleaner like-for-like.
Latency is similar: 282 seconds against 268 seconds for Fable 5.1, so you are not buying cost with slowness.
For general context on how frontier models are ranked for price and quality, see our write-up of the Artificial Analysis Intelligence Index v5.
Does it generalise?
Almost certainly not beyond finance research, and Samaya does not claim it does. The gains are on one benchmark family, in one harness, with training questions written by finance professionals graded the same way as the benchmark. Samaya states that the expert-graded questions used for SFT and offline GRPO are easier than FrontierFinance, and that the model improves on every difficulty third of the benchmark. The hardest third rose from 31.8% to 45.2%.
Open questions the post leaves open:
- Reproducibility: No weights, API or code for Sterling are announced in the text we read.
- Other evaluators: Nobody outside Samaya has published a FrontierFinance run of Sterling.
- Out-of-harness behaviour: It is unclear how the model performs without Samaya's finance tools.
- Frontier models with tuned harnesses: Fable 5.1 was run on the leaderboard configuration; a harness tuned for it might narrow the gap.
What this means for what you build or pay
If you run a repeatable research or analysis workload and have expert-graded examples, this is a template: pick an open model with few active parameters, build or adapt a harness, generate teacher traces from strong models, then use rubric-based reinforcement learning. Samaya's reward-shaping experiment is also useful: a context-length penalty cut cost 30% for 1.1 points of quality at step 20, while a bonus raised quality 0.9 points for 115% more cost. In other words, you can steer the quality and cost trade-off through the reward.
If you do not have expert-graded data, the lesson is different: the expensive part is the grading, not the GPUs. Compare it to NVIDIA's EdgeWise approach of a local small model delegating critical steps to a cloud LLM if you want a cheaper path without training.
For teams choosing among small models today, our small model comparison is the general-purpose baseline; Sterling is the specialist alternative.
What Samaya says comes next
The company lists three directions: multi-agent research with large swarms of smaller agents covering different strategies, firm-specific models that encode a firm's judgement through data, tools and historical decisions, and post-training larger models, where it says early results show even greater gains. It promises a future post on the larger model.
Related reading
- Shopify Sidekick post-trained on Nemotron 3.5 Lightning
- Exowatt fine-tuned Qwen3 vs Claude Opus
- Nemotron 3.5 fine-tuning recipe for IOI and IMO
- Nemotron personas and Lightning LoRA guide
- Small models compared
- Artificial Analysis Intelligence Index v5
- Official: Samaya Sterling post and Samaya's FrontierFinance benchmark page
Details are accurate as of October 28, 2026 and come from Samaya's own post; the benchmark and results have not been independently reproduced.