Samaya AI has launched Sterling, a post-trained model for financial research that scores 58.9% on the FrontierFinance benchmark. This result beats Claude Fable 5.1’s 55.9% while costing $0.61 per question compared to Fable 5.1’s $6.89.
Model specifications and performance
The base engine is NVIDIA’s Nemotron 3.5 Lightning, a mixture-of-experts system with 30 billion total parameters and 3 billion active. The model accepts up to 1 million tokens in context, though Sterling’s specific serving context window is not disclosed.
On the FrontierFinance leaderboard, Sterling outperforms every high-quality system on quality per dollar. It achieves 58.9% overall accuracy and 60.6% in high-effort mode. The hardest third of questions yields a 45.2% score, up from 31.8% for the base model.
System architecture
The team adapted its production finance harness to support a smaller model. System prompts were stripped of detailed guidance, and tool descriptions were merged to reduce length. Two helper models manage context: one condenses long tool results, and another summarises the trace as it grows.
A research budget caps the agent’s working time. Once this limit is reached, a separate report writer generates the final answer based on the gathered findings.
Training methodology
Training occurred on Thinking Machines Lab’s Tinker API using LoRA. The process followed four stages:
- SFT on synthetic traces: The system generated 968 analyst questions. Kimi K3 answered each twice, and an LLM judge selected the superior response. The model learned to call 37 tools with valid arguments.
- SFT on the best of 8 strategies: Kimi K3 answered 1,819 expert-written questions using eight research strategies. The best-graded answer for each question formed the training data. This method met 80% of essential rubric items, compared to 66% for a single best strategy. Filtering reduced the set to 1,157 questions.
- Offline GRPO: GRPO scored all eight teacher answers against one another. No new generation was required, keeping each step inexpensive.
- Online GRPO: The model answered each question eight times and learned from graded attempts. Truncated importance sampling corrected a known mismatch in the MoE sampler.
Benchmark details
FrontierFinance contains 220 public queries scored against expert-written rubrics. The grader employs a three-model judge panel. Samaya states no benchmark question was used in training. Checkpoints were selected based on 100 internal validation questions.
Sterling scores 58.9% at $0.61 and 282 seconds per question. Fable 5.1 scores 55.9% at $6.89 and 268 seconds. Samaya’s own high-effort system scores 56.0% at $1.81.
Each training stage added quality. Scores progressed from 42.5% (base), to 51.4%, 54.8%, 57.5%, and finally 58.9%. Median tool calls rose from 3 to 17. In high-effort mode, Sterling reaches 60.6% at $1.83 and 372 seconds. On the hardest third, accuracy rose from 31.8% to 45.2%.
The model itself costs $0.04 per question. Tools and helper models make up the remaining $0.57 of the $0.61 total.
Comparison table
| Feature | Samaya Sterling | Claude Fable 5.1 | Samaya System, high effort | Nemotron 3.5 Lightning base |
|---|---|---|---|---|
| Total / active params | 30B / 3B | Not disclosed | Not disclosed | 30B / 3B |
| Context | Not disclosed | 1M | Not disclosed | Up to 1M |
| License / access | No public weights or API announced | Proprietary API | Samaya platform | OpenMDW-1.1, open weights |
| FrontierFinance score | 58.9% (60.6% high effort) | 55.9% | 56.0% (58.9 minus 2.9 pp) | 42.5% |
| Cost per question | $0.61 ($1.83 high effort) | $6.89 | $1.81 | Not disclosed |
| Latency per question | 282 s (372 s high effort) | 268 s | 278 s | Not disclosed |
| Token pricing | Not disclosed | $10 / $50 per MTok | Not disclosed | Not disclosed |
FrontierFinance figures for all columns come from the Samaya blog and are self-reported.
What it means
Developers can now train a small model alongside its harness to win in a narrow domain. The gains come from co-training the model and the system rather than relying on a larger base engine. Sterling remains a proprietary system with no public weights or API available at launch.