OpenPipe
Summer 2023AcquiredTurn expensive prompts into cheap fine-tuned models
Pressure-test this opportunity
Turn this teardown into a decision-ready prompt for ChatGPT, Claude, or your agent.
Turn expensive prompts into cheap fine-tuned models
Turn this teardown into a decision-ready prompt for ChatGPT, Claude, or your agent.
If you only have a few minutes to spare, here’s what investors, operators, and founders should know about OpenPipe (S23).
OpenPipe was a developer platform for fine-tuning language models and later training agents with reinforcement learning. Brothers Kyle and David Corbitt founded it in Seattle in 2023, raised a $6.7 million seed, and sold it to CoreWeave in a transaction that closed on September 5, 2025.[1][2]
This was an acquisition success, not a distress story. OpenPipe found that task-specific training could make smaller models cheaper and more reliable, then moved from supervised fine-tuning into GPU-intensive agent reinforcement learning. That direction fit CoreWeave's infrastructure and its newly acquired Weights & Biases platform. Product survival split after the deal: ART remains active, while the original repository is paused and legacy training and inference are being retired.
Kyle and David Corbitt are brothers. Kyle previously worked at Google and led YC's Startup School team; David worked at Palantir and Qualtrics.[3] They began in March 2023 with a browser-control agent that could post to Twitter and book flights. It succeeded about 60% of the time, and each demonstration cost roughly $15 to $20 in input tokens.[4]
The brothers also had task-specific pain. David's Reddit-classification app made each search cost multiple dollars. Kyle's document-translation startup found GPT-3.5 unreliable and GPT-4 too slow.[5] Experiments with FLAN-T5 suggested that a smaller specialized model could approach perfect reliability on tasks where GPT-4 reached roughly 80%, although those founder-reported results were not independently reproduced.[4]
OpenPipe turned that lesson into infrastructure. Developers would prototype with a frontier model, collect prompt and completion pairs from production, train a smaller model, and deploy it behind an OpenAI-compatible interface. The product removed the specialist work of dataset curation, training, feedback, and hosting.
The canonical format requests two verbatim founder quotations. The corpus preserves only one acquisition-era quote suitable for exact reproduction, used later in this report. Other founder findings are paraphrased rather than reconstructed.
OpenPipe's original product was a drop-in OpenAI replacement that logged calls in the background. Production traffic became a candidate dataset without slowing the live response path. Developers filtered logs, imported JSONL, corrected examples manually or with a model, trained open or closed models, evaluated candidates, and deployed through an OpenAI-compatible endpoint.[4]
The Apache-2.0 repository included Python and TypeScript SDKs, request-log filtering, evaluations, hosting, and downloadable weights.[13] Portability was deliberate: customers training Mistral, Mixtral, or Llama models owned their resulting weights and could host elsewhere.[7]
The economic pitch was dramatic but founder-reported. A recipe-classification experiment said a fine-tuned 7-billion-parameter model matched GPT-4 labels 95% of the time and processed more than two million recipes for $19, versus a projected $23,000 with GPT-4.[6] No independent reproduction established dataset selection, label quality, or production generality.
Read the complete post-mortem, the rebuild playbook, and the exact reasons OpenPipe is still worth studying now.