New playbook: when your own AI model pays. Read
Cognition (software, ~300+ employees) reported this on 8 Jul 2026 in company self-report (source). It belongs to the learning loop layer.
SWE-1.7 (July) and SWE-2 (September): open base models (Kimi K2.7 Code, Kimi K3) trained with large asynchronous RL across four datacenters for long-horizon coding in its own agent.
SWE-1.7 FrontierCode 42.3% vs 43.0% for GPT-5.5 (base Kimi 30.1%); SWE-2 reports 50.0%, within one point of Fable 5.1 (50.9%) at 64% lower cost; RL adds 5-6 points on many benchmarks.
The number covers public and Cognition-built coding benchmarks.
“Frontier Intelligence at a Fraction of the Cost,” Cognition said.
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).