Features
- Generates fresh, context-aware test scenarios to avoid benchmark contamination
- Runs a four-stage pipeline: understanding, ideation, rollout and judgment
- Runs multi-turn conversations and tool-use environments with the target model
- Scores results with a judge AI
- Detects sycophancy in AI models
- Evaluates reasoning-heavy models
- Shares reproducible behavior configs as seeds under the Apache-2.0 license
- Integrates with Weights & Biases and Inspect
Pricing Details
💵 Pricing & Plans (February 2026 Status): Plan Price Key Features Open Source $0 (Free) Fully available on GitHub; requires your own API keys for the models you test. Enterprise / Cloud Custom Managed hosting of Bloom; automated CI/CD safety gates; premium support. 🎁 How to Get Started: Go to the Safety-Research/Bloom repository on GitHub. Clone the repo, install the Python dependencies, and write a simple "seed" file describing the behavior you want to test (e.g., "political bias"). Run the command line tool, and Bloom will begin generating and running its interrogation scenarios.