Blackbox

Blackbox presents itself as a high-trust platform for frontier inference. According to the page, prompts are encrypted before leaving the client and decrypted only where the model runs, with zero data retention meaning nothing is kept after the answer. The platform offers access to 300+ models through one endpoint and dedicated deployments of open-weight models. The page also describes secure GPU rental with access controls, spend limits and an audit log, scoped API keys, SSH key checks, and a forward-deployed engineering offering in which agents tune the engine on a customer's workload. It also describes fine-tuning, post-training, and reinforcement learning services available to Enterprise customers. Performance claims on the page include a benchmark comparison of output tokens per second for GPT-OSS-120B on one NVIDIA B200, sourced from Artificial Analysis and Blackbox's own agent run. Enterprise sales contact is the main call to action.

Key features

• End-to-end encrypted inference for every model
• Zero data retention
• Dedicated deployment of open-weight models
• Access to 300+ models through one endpoint
• Model catalog and provider benchmarks
• Scoped API keys
• Spend limits and audit log for GPU deployments
• SSH key checks and no cloud provider keys held by the team
• Secure GPU rental on NVIDIA hardware
• Fine-tuning, post-training, and reinforcement learning services for Enterprise customers
• Forward-deployed engineering for Enterprise customers
• Microsoft Azure deployment option

Best suited for

• Enterprise teams deploying AI models at scale with strict privacy requirements
• Organizations needing encrypted or air-gapped model inference
• Teams wanting to fine-tune or post-train open-weight models on their own data

Pricing

Not stated on the website