3B Activated Model Sterling Beats Claude: 11x Cost Reduction for Financial Research Agents

Samaya AI has released Sterling, a post-trained model designed for financial research scenarios. Built on Nemotron 3.5 Lightning with approximately 3B activated parameters, it scores 58.9% on the FrontierFinance benchmark at a cost of $0.61 per question. For comparison, Claude Fable 5.1 scores 55.9% on the same benchmark at $6.89 per question.

Taken together, these numbers point to more than just a benchmark win—they raise a more interesting question: in vertical domains, can small, specialized models already outperform general-purpose flagship models on both quality and cost simultaneously?

What Sterling Is: A Small Model Post-Trained for Financial Research

Sterling's positioning is very clear—a financial research agent. It is not a fine-tuned version of a general chat model, but a specialized model obtained through post-training on top of Nemotron 3.5 Lightning. With approximately 3B activated parameters, the computational overhead during inference is kept at a fairly low level.

This determines its cost structure. The high unit price of large models largely comes from activating a massive number of parameters on every call, whereas Sterling completes the same type of task with far fewer activated parameters, naturally reducing the compute consumption per inference. The per-question cost gap of $0.61 versus $6.89—roughly 11x—mainly stems from this.

It's worth emphasizing that Sterling targets the specific scenario of financial research—retrieving, reading, and synthesizing financial information to deliver research-oriented answers. Such tasks often involve long inputs and require multi-step reasoning, placing high demands on domain knowledge and information organization, while having very low demand for general capabilities like open-ended chitchat or multilingual creative writing. The advantage of specialized models is precisely reflected in these "narrow but deep" tasks.

Activated parameters is the key concept for understanding Sterling's cost advantage. In a Mixture-of-Experts (MoE) architecture, the total parameter count can be very large, but each forward inference only activates a subset of the expert networks. Activated parameters refer to the scale of parameters actually involved in computation during that inference. It directly determines the floating-point operations per inference, which in turn determines compute cost and latency. By contrast, dense models use all parameters on every inference. Therefore, the same "3B activated" may correspond to a total parameter count far larger than 3B, but inference cost is calculated only on the activated portion. This is also the architectural foundation that has allowed small, efficient models to establish a foothold alongside large models in recent years.

FrontierFinance Benchmark: What Does 58.9% vs. 55.9% Mean

FrontierFinance is a benchmark measuring model performance on financial research tasks. Sterling's 58.9% versus Claude Fable 5.1's 55.9% represents a 3-percentage-point difference. Looking at the absolute values alone, this is not a crushing lead, but when combined with cost, the conclusion becomes entirely different.

In benchmarks, a few points of difference often correspond to whether a model possesses a certain key capability rather than random fluctuation. Sterling, while scoring slightly higher, compresses per-question cost to about one-tenth of its competitor's—which directly changes feasibility in research workflows that require large numbers of model calls: the same budget can run an order of magnitude more queries.

More notably, consider the choice of comparison target. Claude Fable 5.1 belongs to the general flagship tier and is typically regarded as one of the ceilings of overall capability. That a specialized model with 3B activated parameters can surpass it on this vertical benchmark suggests that the combination of "general large model + domain post-training" already has the realistic conditions to replace general flagships on certain tasks.

Benchmark scores need to be understood against two backgrounds: statistics and task composition. On one hand, whether a few percentage points of difference is significant depends on test set size and difficulty distribution; when sample size is insufficient, a small lead may fall within the confidence interval. On the other hand, benchmarks are usually weighted combinations of multiple subtasks, and a close total score does not mean consistent performance across subtasks—one side may lead on retrieval-type questions while the other is stronger on numerical reasoning. Therefore, 58.9% vs. 55.9% is better interpreted as "a slight lead within the same performance band," and its true significance should be evaluated jointly with the cost dimension rather than viewed in isolation as a leap in capability tier.

Why the Post-Training Route Pays Off More in Vertical Scenarios

Sterling chose to post-train on Nemotron 3.5 Lightning rather than train from scratch or directly call closed-source large model APIs. This route has several practical advantages.

First is data efficiency. The financial research domain has abundant specialized corpora, research reports, and financial statement texts, and post-training allows the model to fully absorb these distributions without redundant investment in general capabilities. Second is cost controllability. The base model itself is not large in parameter scale, and the compute requirements for both post-training and inference are within acceptable ranges, keeping per-call pricing very low. Third is controllability. A self-owned post-trained model is easier to adjust as needed for data compliance, output style, and domain terminology handling—important for heavily regulated industries like finance.

The trade-offs are equally real. Specialized models will likely underperform general flagships on non-financial tasks, with limited generalization. If a team's needs span multiple domains, a single specialized model may not suffice, possibly requiring a combination of multiple models or retaining a general model as a fallback.

Post-training refers to the process of shaping general capabilities into task-specific capabilities on top of a pretrained base model through supervised fine-tuning, preference alignment, continued training on domain data, and similar methods. Its difference from "pretraining from scratch" is that the pretraining stage has already consumed the vast majority of compute and general corpora, so post-training requires only a relatively small amount of domain data to significantly change model behavior. For scenarios like financial research, where data is highly specialized and task boundaries are clear, post-training offers a very high return on investment—no need to rebuild the model, just to make it "understand" the distributions of financial statements, research reports, and regulatory texts. This is the fundamental reason Sterling can quickly reach specialized-level performance on a small base.

Practical Impact on Financial AI Applications

An 11x cost difference will directly change the design approach for financial research agents. In the past, when building research assistants with general flagship models, cost often became the bottleneck for scaling—a single deep research task might require dozens or even hundreds of model calls, and high unit prices constrained both product pricing and usage frequency.

Once models like Sterling bring per-question cost down to the $0.61 level, agents can conduct multi-round retrieval, cross-validation, and information synthesis more frequently without meticulously optimizing every step. This is especially critical for financial research tasks requiring long-chain reasoning, because research quality often depends on the model being willing to "check a few more times."

Another impact is deployment choice. Specialized small models can be self-hosted, keeping data within the internal network—a hard requirement for institutions handling non-public financial information and investment strategies. General API solutions often fail compliance checks in such scenarios.

What to View with Caution

Sterling's release information currently focuses on benchmark scores and cost, with limited verifiable details. As a benchmark, FrontierFinance's coverage and task difficulty distribution are not elaborated in the source material, making it difficult for outsiders to judge which capabilities the 58.9% score specifically corresponds to and where it loses points on subtasks.

Cost figures also need to be understood in context of usage. $0.61 is the per-question cost; total cost in actual products also depends on number of calls, context length, and engineering overhead. Additionally, the effectiveness of specialized models is highly dependent on the quality and coverage of post-training data, and how they perform on real-world financial problems beyond the benchmark requires more independent evaluation to verify.

General flagship models will not stand still either. There will be new versions after Claude Fable 5.1, and general models' performance in vertical domains is also improving. Whether Sterling's current lead can be maintained depends on the pace of Samaya AI's subsequent data iteration.

Summary

Sterling's value lies not in the 58.9% score itself, but in the clear path it demonstrates: using a medium-scale base with domain post-training to match general flagships on specific tasks at an order-of-magnitude lower cost. For scenarios like financial research, where task boundaries are clear and data is highly specialized, the cost-effectiveness of this path is already hard to ignore. What's worth watching next is independent reproduction results and how this type of specialized model replicates in more vertical domains.