Samaya AI launches Sterling agent, claims 11x cheaper finance research than Claude Fable 5.1
Samaya AI released Sterling, a 3B-active-parameter agent built on NVIDIA’s Nemotron 3.5 Lightning, optimized for financial research. The model scores 58.9% on Samaya’s FrontierFinance benchmark, surpassing Claude Fable 5.1’s 55.9% while costing $0.61 per query versus $6.89. Sterling’s performance stems from co-training the model with its harness, which includes a condenser for long tool results,…
Key points
- Sterling scores 58.9% on FrontierFinance, 3 points above Claude Fable 5.1’s 55.9%, at 11x lower cost per query
- Built on NVIDIA’s 30B Nemotron 3.5 Lightning with 3B active parameters, trained via LoRA and GRPO in four stages
- Harness includes context tools (condenser, compactor) and a research budget cap to manage agent efficiency
The model runs on a single H100 80GB GPU in BF16 with up to 1M-token context, though Sterling’s serving context window is undisclosed. Samaya reports Sterling beats every system on FrontierFinance in cost-per-quality, with a median of 17 tool calls per query. High-effort mode achieves 60.6% at $1.83 per question. The company emphasizes its approach—small model, efficient harness—as a scalable recipe for domain-specific agents. All benchmark results are self-reported, and no training data overlaps with FrontierFinance’s 220 public queries.
Model pages: Claude Fable 5.1 → · Kimi K3 →
Samaya AI Releases Sterling: A 3B-Active Finance Research Agent Model that Beats Claude Fable 5.1 on FrontierFinance at 11x Lower Cost
MarkTechPost · 30 October 2026
Loading the full article…
This text was published by MarkTechPost and written by Asif Razzaq. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Agents & Tools
All →- Andy Jassy says outside shopping agents are not yet great or convenient · 3 src
- Claude outperforms Roblox Studio’s built-in AI in three game tasks · 5 src
- Anthropic study finds experts follow AI agent's hidden preference 86% of the time · 7 src
- Anthropic launches Claude Duet, a proactive AI assistant for US paid users · 3 src
- Google shuts down AI coding agent Jules · 1 src
Comments
via GitHub Discussions