Last updated: 2026-10-13
OpenPipe's hosted fine-tuning platform is inactive as of 13 October 2026. The company said in May 2026 that training and inference on OpenPipe.ai would end on 30 July 2026, after CoreWeave rebuilt the features inside Weights & Biases. The open-source ART trainer is still maintained. HokAI keeps this page for reference.
About OpenPipe
OpenPipe was a fine-tuning and reinforcement learning platform founded in 2023 and acquired by CoreWeave in September 2025. It captured production LLM calls, curated them into datasets and trained cheaper open models to replace frontier-model calls on narrow tasks.
That hosted product is gone. In a notice dated 18 May 2026, the founders said the platform's features had been rebuilt inside Weights & Biases and that training and inference on OpenPipe.ai would stop on 30 July 2026. The sign-in app cannot be reached, and the pricing page now carries only a short pointer to the migration. Fine-tuning continues on CoreWeave as Serverless SFT, which CoreWeave's documentation lists as a public preview.
What survives is ART, the Agent Reinforcement Trainer. It is open source under Apache-2.0 on GitHub, was still receiving commits on 12 October 2026, and is the recommended way to call CoreWeave's serverless training API.
Who should look elsewhere: anyone who wants a self-contained fine-tuning dashboard with evaluations and data pipelines, because those features were not carried over. For hosted serving of open-weight models, compare Together AI and Fireworks AI; for running your own training jobs, see Modal and Hugging Face. Check each page for whether it covers fine-tuning or only inference.
Pricing
The hosted OpenPipe platform no longer sells anything: training and inference ended in July 2026, per the vendor migration notice. Earlier per-token rates are not shown because they cannot be verified any more. ART is free to use; hosted training now runs on CoreWeave Serverless SFT, which is in public preview.
Key Features
- ART trainer (open source): Agent Reinforcement Trainer is a GitHub library that trains multi-step agents with reinforcement learning, and it keeps receiving commits.
- Serverless SFT through ART: CoreWeave's Serverless SFT is the successor path and can be driven from ART or the Serverless Training API.
- Migration to W&B Inference: Customers could move existing models to W&B Inference with an in-app tool before the shutdown.
Pros
- ART stays open source, so the training code outlives the hosted product.
- The vendor gave a documented migration path and a dated shutdown, rather than an abrupt closure.
Cons
- Nothing is left to subscribe to: the hosted service closed to new training and inference jobs in July 2026.
- Several lesser-used features, including DPO and the relabeling workflows, were not carried over to the successor.
- Serverless SFT is a public preview product, so limits and pricing may change.
Frequently Asked Questions
Is OpenPipe still active in 2026?
No. As of 2026-10-13, the hosted OpenPipe platform is no longer operational. The vendor's migration notice dated 18 May 2026 says training new models and running inference on OpenPipe.ai ended in July 2026, and the app.openpipe.ai host no longer resolves. HokAI marked this page inactive after checking both.
What replaced the OpenPipe platform?
CoreWeave rebuilt the core features inside Weights & Biases. Its Serverless SFT product, listed as public preview in CoreWeave's documentation, fits a LoRA adapter to curated input and output pairs, then serves every checkpoint on Serverless Inference. You can run it through OpenPipe's ART framework or the Serverless Training API.
Is the OpenPipe ART library still available?
Yes. ART lives at github.com/OpenPipe/ART under a permissive open source licence, is not archived, and had a commit pushed this month. The OpenPipe site says the team continues to support it.
What happened to existing OpenPipe models and data?
OpenPipe offered an in-app migration tool that moved existing models to W&B Inference before the deadline. The notice listed a set of lesser-used features that would be sunset with the platform, so any data tied to them had to be exported before the deadline.
Where can you fine-tune models now?
CoreWeave Serverless SFT is the direct successor. HokAI also lists Together AI, Fireworks AI, Modal and Hugging Face in nearby categories; open each page to see whether it covers training or only serving before you choose.
HokAI guides covering OpenPipe
- Hugging Face vs OpenPipe: Which Should You Use in 2026?: OpenPipe stopped training on 30 July 2026. We compare Hugging Face TRL, CoreWeave Forge, Together AI, Fireworks and Overmind on price, RL support and data.