New playbook: when your own AI model pays. Read
Kasisto, deployed at 40+ financial institutions (banking software, ~2,000 employees) reported this on 30 Jul 2026 in peer-reviewed paper (source). It belongs to the learning loop layer.
FinRAG-12B, a 12B model fine-tuned for grounded banking Q&A with citations and trained to say 'I don't know' (22% of training examples unanswerable), serving customer queries in production.
Answer quality 6.21 vs 5.72 for GPT-4.1 on banking QA; ~$0.001 per query, 20-50x cheaper and 3-5x faster than GPT-4.1; training cost ~$1,800; deployed at 40+ financial institutions.
The number covers offline evaluation on 258 banking QA examples from 3 institutions plus serving-cost measurement; baseline is GPT-4.1, an older frontier model.
“…the model delivers 3–5x faster responses at 20–50x lower cost compared to GPT-4.1,” Kasisto, deployed at 40+ financial institutions said.
Reviewed corrections are kept as test cases and later used to train models.
Read the guide to the learning loop.
96%
lower estimated serving cost for one agent
A daily continual-learning loop for Sidekick's GraphQL agent: an expert-calibrated LLM judge scores production conversations, a self-healing pipeline repairs failures, and the repaired trajectories are trained into a smaller model (SFT, then GRPO against the judge) that replaces the frontier model (source).
13.8x
lower cost per task, with higher accuracy than the best frontier model
A Qwen3-235B model fine-tuned on Tinker with investor-labelled data to do six daily information-filtering tasks (e.g. which news matters to a macro investor) (source).
1% → 99%
valid structured output after fine-tuning a small model
A proof of value on Red Hat OpenShift AI: GPT-4.1 generated synthetic training data, then small models (Phi-4 Mini 3.8B, Qwen 2.5 1.5B/7B) were fine-tuned for multilingual structured extraction from social media, deployable locally (source).