A bench that runs the same banking questions through several hosted models and reports accuracy, cost, and speed side by side, on real data.
| Model | Accuracy, Fine | Accuracy, Coarse | Cost / 1k Messages | p50 Latency | p95 Latency | Brier Score |
|---|---|---|---|---|---|---|
| Amazon Nova Lite | 70.5% | 86.4% | $0.0492 | 498 ms | 3,308 ms | 0.2609 |
| Meta Llama 3.3 70B | 72.2% | 86.3% | $0.4025 | 1,417 ms | 9,465 ms | 0.2316 |
| Mistral Large 2402 | 69.3% | 84.1% | $0.4404 | 1,753 ms | 10,028 ms | 0.2403 |
Full Banking77 test split, 3,080 messages, run once per model against AWS Bedrock (region us-east-1, prices as of 2026-09-16). From results.json in the repo below.
Three hosted Bedrock models are in this run: Amazon Nova Lite, Meta Llama 3.3 70B, and Mistral Large 2402. Two Claude models are priced in the pipeline (data/prices.json) but blocked at the AWS account level today and are not included here.
Type a banking-support message and send it to all three models.
Live runs switch on when the hosted keys are added.
The bench, the prompt, and the scoring pipeline: github.com/alphan-ml/model-bench