OpenPipe

What is OpenPipe?

OpenPipe is a model-training, evaluation, and hosting platform for teams building LLM applications and agents. Its SDK records requests and responses from an application, lets teams filter or import that data into datasets, and uses the resulting examples to train specialized models. The product supports supervised fine-tuning, direct preference optimization, evaluations, and hosted deployments.

OpenPipe implements an OpenAI-compatible inference format, so applications can move from a third-party base model to a trained deployment with limited integration changes. The platform is intended for repeated production tasks where a smaller specialized model may reduce latency or cost while preserving the behavior measured by the team's evaluations.

How to Use OpenPipe

1. Integrate the SDK into an existing model-powered application.
2. Capture representative requests, responses, feedback, and useful metadata.
3. Filter logs or import approved examples to create a training dataset.
4. Train a model with fine-tuning or preference optimization and compare it against baselines.
5. Deploy the selected model, monitor its outputs, and continue improving the dataset.

Core Features

• Request logging: Captures application inputs and outputs with tags for later analysis.
• Dataset management: Builds datasets from logs or imported OpenAI-compatible JSONL files.
• Fine-tuning and DPO: Trains specialized models from examples and preference data.
• Evaluations and criteria: Compares models and measures behavior against defined expectations.
• Hosted inference: Deploys trained models behind an OpenAI-compatible chat interface.
• Caching and fallback: Supports response caching and fallback behavior for production workloads.

Use Cases

• Agent reinforcement and optimization: Improve a task agent from collected trajectories, outputs, and preferences.
• Cost reduction: Replace expensive repeated prompts with a smaller model trained for a narrow task.
• Latency improvement: Host a specialized model suited to a high-volume workflow.
• Model evaluation: Compare trained, base, and third-party models on the same inputs and criteria.

Pricing

Training is billed by model size and dataset tokens; current documentation lists rates from $0.48 per million training tokens for models up to 8B parameters to $2.90 for models of 70B parameters or larger. Hosted inference is billed either per token or by hourly compute units, depending on the model. Third-party fine-tuned models are billed directly by their providers without an additional OpenPipe markup. Enterprise plans with volume discounts, on-premises deployment, support, SLAs, security, and expanded storage use custom pricing.

Frequently Asked Questions

Does OpenPipe require changing an application's response parser?

Its trained deployments implement an OpenAI-compatible inference format, which reduces the integration changes needed when switching models.

What data is used for training?

Teams can use captured request logs, tagged examples, imported JSONL data, and preference information. They should remove sensitive or unauthorized data before building a dataset.

Does a fine-tuned model automatically outperform the original model?

No. Performance depends on task definition, data quality, training configuration, and evaluation design. Teams should compare models on representative held-out examples before deployment.