Sterling finance research agent outperforms Claude Fable 5.1

What changed

Samaya AI released Sterling, a specialized finance research agent built on NVIDIA’s Nemotron 3.5 Lightning, a 30B mixture-of-experts model with 3B active parameters. On Samaya’s FrontierFinance benchmark, Sterling achieves 58.9% accuracy, surpassing Claude Fable 5.1’s 55.9%. The cost difference is substantial: Sterling operates at $0.61 per question versus $6.89 for Fable 5.1, representing an 11-fold cost reduction.

Sterling was developed through a four-stage training process using LoRA on Tinker, a training API from Thinking Machines Lab. The first supervised fine-tuning stage trained the model on 968 synthetic analyst questions. For these questions, the teacher model Kimi K3 generated two responses for each question, and an LLM judge selected the stronger response. This stage taught Sterling to call 37 financial tools with valid parameters and gather evidence before reporting findings.

The second supervised fine-tuning stage focused on research strategy selection. Samaya prompted Kimi K3 to answer 1,819 expert-written questions using eight different research strategies. Each answer was graded against expert-written rubrics, and the best-performing answer became training data. This approach proved effective: selecting the best of eight strategies satisfied 80% of essential rubric items on average, compared to approximately 66% for a single strategy. After filtering by length and tool call patterns, 1,157 questions remained in the final training set.

The third stage employed offline group relative policy optimization (GRPO), which scored the eight existing teacher responses against each other without requiring new generation. This efficient approach allowed the model to learn how different research strategies perform on the same questions.

The fourth stage used online GRPO, where Sterling generated eight responses per question and learned from graded attempts. This stage refined the model’s ability to improve through feedback. Across all four stages, performance climbed from 42.5% to 58.9%.

Sterling’s architecture includes careful harness design to support a small model. Detailed guidance was removed from system prompts and tool descriptions, allowing the model to learn these through training rather than explicit instruction. Two helper models manage context constraints: a condenser compresses long tool results, and a compactor summarizes growing research traces. A research budget caps the investigation duration, after which a separate report writer generates final output from accumulated findings. This design allows the model to sustain research across many tool calls despite having only 3B active parameters. The base model fits on a single H100 80GB GPU in BF16 format with 256K context.

Why it matters

Sterling’s results reshape the cost-benefit calculation for deploying AI in finance workflows. Organizations previously needed frontier models to achieve acceptable research quality; Sterling demonstrates that smaller, specialized models can match or exceed those results at a fraction of the cost. This finding extends beyond finance, suggesting that targeted co-training of models with their operational harnesses can be more effective than scaling up model size alone.

The training recipe employed here validates an approach that combines deliberate system design with focused post-training. Rather than expecting a smaller model to perform well without infrastructure support, Samaya built financial tools, context compression, and research planning directly into the system, then trained the model to operate effectively within these constraints. This differs from the common pattern of attempting to solve capability gaps through larger models.

The results are particularly significant for cost-conscious organizations evaluating AI infrastructure decisions. An 11x cost reduction with higher accuracy on a relevant benchmark could shift procurement choices away from frontier APIs toward specialized alternatives or self-hosted solutions.

However, all metrics come from Samaya’s own benchmark and testing environment. The FrontierFinance leaderboard contains 220 public queries scored against expert-written rubrics, but independent validation from other organizations would strengthen confidence in the results. Additionally, performance varies with question difficulty: Sterling reaches 60.6% in high-effort mode but drops to 45.2% on the hardest third of questions.

What to test

Before deploying Sterling for production financial research, evaluate these dimensions:

Benchmark alignment with your workflows. Verify that FrontierFinance questions reflect actual research tasks you need to solve. The benchmark covers questions such as interest coverage ratios and debt service capability analysis. Run a sample of your own questions through Sterling and a baseline model to determine whether benchmark results translate to your use cases.

Real-world cost measurement. The $0.61 figure includes the model, financial tools, and helper models. Measure actual costs in your environment, accounting for tool API calls, infrastructure overhead, and the report writer’s resource consumption. Question difficulty, tool call patterns, and timeout frequency may shift per-question costs.

Quality consistency across difficulty levels. Sterling’s average performance is 58.9%, but the hardest third of questions drops to 45.2%. Identify your mission-critical research questions and test them directly. Understand what failure modes occur in the lower performance quartile and whether they are acceptable for your use case.

Tool coverage requirements. Sterling accesses 37 financial tools integrated into Samaya’s harness. Audit whether your required data sources are included. For missing tools, understand the integration effort and potential impact on model training and inference speed.

Latency tolerance. Sterling requires 282 seconds per question compared to 268 seconds for Fable 5.1. Assess whether this latency is acceptable for your workflows. Verify that a single H100 meets your throughput requirements or whether you need to distribute across multiple GPUs.

Independent validation. All published results are self-reported on Samaya’s own benchmark with Samaya’s test set. Request or conduct independent evaluation against public benchmarks or your own held-out data. Run head-to-head comparisons with Fable 5.1 and other agents in your environment.

Common questions

How does Sterling sustain long research workflows with limited parameters?

Sterling uses two helper models to manage context. A condenser compresses lengthy tool results into shorter summaries, while a compactor periodically summarizes the growing research trace. A research budget caps investigation duration, triggering the report writer to generate final output once the limit is reached. This approach allows even a 3B-active model to maintain coherence across dozens of tool calls.

What happens when Sterling cannot find sufficient evidence?

Once the research budget expires, the report writer generates output from available evidence, potentially creating partial answers. Samaya’s documentation does not specify how frequently insufficient evidence occurs or how the model signals low-confidence findings.

How is Sterling currently available?

Samaya’s announcement does not specify access methods, licensing terms, or whether model weights are open source. Check Samaya’s website or documentation for availability and pricing details.

Can Sterling be adapted to other domains beyond finance?

The post-training recipe uses finance-specific data, rubrics, and tools, so domain adaptation would require domain-specific training data, expert-written rubrics, and tool harnesses. While the four-stage training approach may be portable, equivalent performance in other verticals is not guaranteed without validation.

The conclusion

Sterling demonstrates that coordinated design of models and their operational harnesses can outperform larger frontier models on specialized tasks while dramatically reducing costs. On Samaya’s FrontierFinance benchmark, Sterling achieves 58.9% accuracy at $0.61 per question, beating Fable 5.1’s 55.9% at $6.89 per question.

The key insight is not simply that smaller models can work in narrow domains, but that systematic co-training of model and harness yields compounding gains. Each training stage built on the previous one, improving from 42.5% baseline to the final 58.9%, while the harness design ensured the model could effectively use financial tools and manage context constraints.

Before adopting Sterling, validate performance on your actual workflows and measure real-world costs in your environment. The performance gap between Sterling and Fable 5.1 is modest on average, and the 45.2% accuracy on the hardest questions warrants investigation. Independent evaluation would also increase confidence in the results.

Watch for independent benchmarks testing Sterling against other agents and evidence of how results translate to real analyst workflows. If the training techniques prove transferable, similar cost-quality tradeoffs could emerge in other specialized domains. For immediate adoption, the most actionable takeaway is the training methodology: small models can compete effectively in narrow domains when the model and its operational system are deliberately co-trained.

Related reading

• GPT-6.1 Terra vs Claude Opus 5.5: Pricing and Performance
• GPT-6.1 Sol and Claude Opus 5.5: 75% price cuts reshape LLM market
• AKV cache engineering cuts agent costs 27.7%