Samaya Sterling: 3B-Active Post-Trained Model Beats Claude Fable 5.1 on Finance Benchmark at 11x Lower Cost
A 30B mixture-of-experts model with 3 billion active parameters has beaten Claude Fable 5.1 on a financial research benchmark at one-eleventh the cost.
Samaya AI published results for Samaya Sterling on October 28, 2026. Sterling runs on NVIDIA’s open-source Nemotron 3.5 Lightning (30B total parameters, 3B active, reasoning enabled), post-trained over four stages using LoRA on Thinking Machines Lab’s Tinker API. It starts from a base accuracy of 42.5% on Samaya’s 220-question FrontierFinance benchmark and reaches 58.9%, 3.0 percentage points ahead of Claude Fable 5.1’s 55.9%.
The Numbers
| System | Score | Cost / Question | Latency |
|---|---|---|---|
| Samaya Sterling | 58.9% | $0.61 | 282s |
| Legacy Samaya System (high) | 56.0% | $1.81 | 278s |
| Claude Fable 5.1 (Finance Agent v2) | 55.9% | $6.89 | 268s |
| Muse Spark Open (Finance Agent v2) | 53.7% | $1.38 | 234s |
| Claude Opus 5.5 (Finance Agent v2) | 50.9% | $1.87 | 148s |
The cost comparison is system-to-system, not model-to-model: Sterling uses Samaya’s proprietary harness and co-designed tools, while Fable 5.1 runs under the open Finance Agent v2 harness. The research overhead (tools, condensation, judge models) is included in both cost figures.
How It Was Built
Training ran in four stages: two supervised fine-tuning rounds where the model learned financial research workflows from eight diverse teacher models, followed by offline RL and then online RL against live mistakes. The post-training was joint: model and harness evolved together, with the harness providing high-quality financial tools the model learned to use through training.
The result shows in tool-call depth. Median tool calls per question rose from 3 (base) to 17 (Sterling). Sterling also improves across all three difficulty thirds of the benchmark, including the hardest: from 31.8% to 45.2% at the top of the difficulty range.
In high-effort mode, Sterling reaches 60.6% at $1.83 per question, still well below Fable 5.1’s $6.89.
What This Shows
The headline is less about Nemotron Lightning itself and more about what co-training does to small models in constrained domains. Sterling’s NVIDIA base model is freely available; the gains come from four training stages adapted to a specific workflow. Samaya notes that early results with a larger model on the same recipe show greater gains, suggesting the technique scales up, not just down.
FrontierFinance covers six use cases across the investment process: discovery, research, financial modeling, earnings analysis, screening, and macro research. The benchmark uses 220 public questions with expert-written rubrics averaging 52.5 items per query, graded by a 3-model judge panel. No training questions were drawn from the benchmark; checkpoint selection used 100 internal validation questions.
FrontierFinance’s current ceiling is 58.9%. Frontier models (Opus 5.5 at 50.9%, GPT-5.6 Sol at 46.8%) run on the standard Finance Agent v2 harness without task-specific post-training and are priced accordingly. Sterling’s win depends on the co-adapted harness, but that harness is available to Samaya customers, making the cost-quality curve a product reality, not just a lab result.