Samaya AI Releases Sterling: A 3B-Active Finance Research Agent Model that Beats Claude Fable 5.1 on FrontierFinance at 11x Lower Cost

Samaya AI says its new Sterling model, a LoRA post-train of NVIDIA's Nemotron 3.5 Lightning with 3B active parameters out of 30B, scored 58.9% on its FrontierFinance benchmark. Claude Fable 5.1 scored 55.9%. Sterling's cost per question was $0.61 against $6.89, roughly an eleventh of the price, at similar latency.

The caveat sits right next to the headline: FrontierFinance is Samaya's own benchmark, and every number here is self-reported. Still, if the result holds up, it is a sign that narrow, expensive research work in finance does not have to run on the biggest model money can buy.

Our read is that the model is not the main story here. The system around it is. Samaya cut guidance out of the system prompt, merged tools, added a condenser for long tool results and a compactor for the growing trace, capped the research budget, and handed the final write-up to a separate report writer. Then it trained the small model inside that setup. That is a practical lesson for anyone building agents: design the harness and the model together, and a small model can punch well above its size in one domain.

The training ladder is honest about where the gains came from. The base model scored 42.5%, and each of the four stages added points, ending at 58.9%. Median tool calls climbed from 3 to 17. In other words, the model learned to do more research before answering, which is what you want from an analyst. On the hardest third of questions it jumped from 31.8% to 45.2%. That is real progress, and it also means it still misses more than half of the hard ones.

The cost breakdown is the part I keep coming back to. The model itself costs about $0.04 per question. Tools and helper models make up the rest. So the price fight in agentic AI is moving from tokens to everything wrapped around them, and teams that only compare per-token pricing will miss where the money actually goes.

Now the eyes-open part. A vendor topping its own leaderboard is the weakest kind of evidence we get. Samaya says no benchmark question was used in training and that checkpoints were picked on a separate internal validation set, which is good practice, but nobody outside has reproduced this yet. And Sterling has no public weights or API announced, so for now it is a capability inside Samaya's platform, not something the broader market can test.

The bigger risk is cultural. Finance research that is cheap enough to run on every question gets run on every question. If a 45% hit rate on hard queries turns into memos that look polished and get skimmed, errors move faster and look more confident. Cheap analysis is useful only if review keeps up.

Sterling was trained with LoRA on Thinking Machines Lab's Tinker training API, using Kimi K3 as a teacher across two supervised stages and two GRPO reinforcement stages. FrontierFinance has 220 public queries graded against expert-written rubrics by a three-model judge panel.