Samaya Says Its Small Sterling Model Beats Claude Fable 5.1 at Finance Research for a Fraction of the Cost
In self-reported results on its own FrontierFinance benchmark, Samaya's 3B-active-parameter model scores 58.9% to Fable 5.1's 55.9%, at $0.61 per question versus $6.89.
Published
Last updated / Fact-checked on
Samaya AI says a small post-trained model can beat a frontier model at financial research. Its Samaya Sterling scores 58.9% on the company's FrontierFinance benchmark, against 55.9% for Claude Fable 5.1, at about one-eleventh the cost per question. Every result here is self-reported by Samaya on a benchmark it built.
Photo: Goumbik / Pixabay
The claim at a glance
| Metric | Samaya Sterling | Claude Fable 5.1 |
|---|---|---|
| FrontierFinance score (self-reported) | 58.9% (60.6% high-effort) | 55.9% |
| Cost per question | $0.61 ($1.83 high-effort) | $6.89 |
| Latency per question | 282 s (372 s high-effort) | 268 s |
| Parameters | 30B total, 3B active (MoE) | Not disclosed |
| Access | No public weights or API announced | Proprietary API |
Samaya published the results in an October 28 blog post. MarkTechPost covered them on October 30 and noted that all the figures are self-reported.
What Sterling is
Sterling is NVIDIA's open Nemotron 3.5 Lightning model, a 30-billion-parameter mixture-of-experts model with 3 billion parameters active at a time, post-trained with LoRA on Thinking Machines Lab's Tinker training API. Sterling works as a research agent. It plans, calls finance tools, gathers evidence and hands its findings to a separate report writer.
Samaya says the harness, the software around the model, mattered as much as training. It took detailed guidance out of the system prompt and tool descriptions, merged tools to shorten the prompt, and added two helper models: a condenser that shortens long tool results and a compactor that summarizes the growing trace. A research budget caps how long the agent works.
How it was trained
Samaya describes four stages:
- Synthetic SFT. The teacher model, Kimi K3, answered 968 synthetic analyst questions twice each, and an LLM judge kept the better answer. Sterling learned to call 37 tools with valid arguments.
- Best-of-eight SFT. Kimi K3 answered 1,819 expert-written questions using eight research strategies, and the best-graded answer for each became training data. After filtering, 1,157 questions remained.
- Offline GRPO. Reinforcement learning compared all eight teacher answers against one another, without generating anything new.
- Online GRPO. Sterling answered each question eight times and learned from the graded attempts. Truncated importance sampling corrected a known sampling mismatch in mixture-of-experts models.
Scores rose at every stage: 42.5% for the base model, then 51.4%, 54.8%, 57.5% and 58.9%. The median number of tool calls per question went from 3 to 17. On the hardest third of questions, the score climbed from 31.8% to 45.2%.
Caveats on the benchmark
FrontierFinance is Samaya's own benchmark. It has 220 public questions graded against rubrics written by experts. Samaya says no benchmark question was used in training and that it picked checkpoints on a separate set of 100 internal validation questions. Independent replication has not been published.
The cost comparison also depends on the setup. Samaya says the model itself costs $0.04 per question, while tools and helper models make up the rest of the $0.61. The Fable 5.1 figure comes from running that model in Samaya's evaluation. MarkTechPost lists Fable 5.1's token pricing at $10 per million input tokens and $50 per million output tokens.
Why it matters
The broader argument is that training a small model together with its harness can win in a narrow field. Samaya says early results from a larger model trained the same way show bigger gains and promises a future post. Sterling has no public weights or API yet, so outsiders can't test it.
What to watch
- Independent evaluations of Sterling on finance benchmarks Samaya didn't create.
- Availability. Watch for a public API or weights.
- The larger model Samaya says is in training.
More coverage is on our Artificial Intelligence hub.
How we reported this
We read Samaya's technical blog post and MarkTechPost's write-up in full and checked every number against both. Where they differed in detail, we used the blog as the primary source. We label all benchmark results as self-reported because Samaya built FrontierFinance and ran the evaluations itself.
Published and fact-checked October 31, 2026.
Sources
- Samaya, "Samaya Sterling: A post-trained model for state-of-the-art financial research," Oct. 28, 2026: https://samaya.ai/blog/rl-post-trained-agentic-model
- MarkTechPost, Asif Razzaq, "Samaya AI Releases Sterling: A 3B-Active Finance Research Agent Model that Beats Claude Fable 5.1 on FrontierFinance at 11x Lower Cost," Oct. 30, 2026: https://www.marktechpost.com/2026/10/30/samaya-sterling-post-trained-finance-research-model-frontierfinance/
DailyMech used AI tools to help research and draft this article; it was fact-checked by our editors.