New playbook: when your own AI model pays. Read
Sundial (software, <50 employees) reported this on 16 Sep 2026 in company self-report (source). It belongs to the learning loop layer.
TeXtinguisher, a fine-tuned Inkling-Small (SFT then GRPO on Tinker) that fixes LaTeX compile errors inside the editor, trained on 3,978 real error/fix pairs mined from TeX.StackExchange.
85.2% of fixes compile vs 80.7% for GPT-5.5 and 82.4% for Claude Fable 5.1, at 0.95 s median vs 2.47-5.72 s; within six points of GPT-6 Astra at a fourteenth of its cost.
The number covers 647 held-out real single-file errors.
Reviewed corrections are kept as test cases and later used to train models.
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).