OpenPipe

OpenPipe is a developer platform for creating task-specific large language models and AI agents. Its original product captured a customer's existing prompts and completions from their current LLM provider through a reporting SDK, then used those prompt-completion pairs as training data to fine-tune a smaller, specialized model that could be used as a drop-in replacement for the original prompt. Developers could kick off a training job from the company's UI and then either self-host the resulting model or have OpenPipe host it, without changing application code or prompts. The company positioned fine-tuning as a way to cut inference cost and latency while maintaining or improving output quality, targeting full-stack and AI engineers who lack dedicated data-science or machine-learning specialists.

The company later focused on reinforcement learning (RL) for agent training, describing itself as a platform for building AI agents that learn from experience and improve in accuracy, performance and reliability over time. It maintains Agent Reinforcement Trainer (ART), an open-source library for training agents with RL, and also offers tooling that automatically adapts AI models to novel tasks so enterprises can obtain custom models without managing the training process themselves.

Founding story

Co-founders Kyle and David Corbitt each encountered the cost and latency limits of general-purpose models before starting the company. David built an application that searched Reddit and classified results along user-specific dimensions, where each search cost multiple dollars with GPT-3.5; Kyle built and sold a startup involving official document translation, where GPT-3.5 could not reliably follow translation requirements and GPT-4 was too slow for a good user experience. The pair had earlier written a web agent running on GPT-4 (Taxy.AI). Kyle previously founded the family-history startup Emberall and worked as an engineer at Google and at Y Combinator, where he led the Startup School team; David was previously an engineer at Qualtrics and Palantir and co-founded the video legacy startup GenerationalStory.

Business model

OpenPipe sold a hosted developer tool: customers integrated its SDK into an existing codebase, trained specialized models on the platform, and then ran those models either self-hosted or hosted by OpenPipe. Alongside the commercial platform it maintained open-source tooling (the ART library) used by developers to train agents with reinforcement learning.

Paid access to the platform for training and hosting fine-tuned models and agents; the sources do not detail pricing.

Traction

As of March 2024 the company said it had saved customers over $3 million in inference costs since September 2023. In September 2025 CEO Kyle Corbitt stated that over the prior two years OpenPipe had helped hundreds of customers move to smaller, faster models, saving them millions of dollars. Team size was listed as 2 on its Y Combinator profile and reported as 10 employees at the time of the acquisition.

Latest developments

CoreWeave announced on September 3, 2025 a definitive agreement to acquire OpenPipe, citing reinforcement learning and fine-tuning capabilities that deepen vertical integration of its stack following its earlier acquisition of Weights & Biases. Terms were not disclosed. OpenPipe's employees join CoreWeave and its customers become CoreWeave customers.