Pi v0.80.6 through 4.5 expanded model support, improved authentication, and added powerful new capabilities for managing local models and controlling AI reasoning.
Models and Reasoning
- Extended thinking levels: New maximum thinking level above xhigh, natively supported on GPT-5.6 and adaptive Claude models, available across CLI, SDK, and RPC.
- Claude Opus 5: Now available on Anthropic, Amazon Bedrock, and GitHub Copilot with adaptive thinking modes and 1M context window.
- Constrained sampling: Strict JSON Schema and grammar variants now supported across OpenAI, Anthropic, Amazon Bedrock, Google Gemini, and Mistral.
- Grok 4.5 thinking levels: xAI Grok now supports three thinking levels for responses via device-code OAuth.
Model Management and Providers
- Local llama.cpp integration: Connect to llama.cpp routers, search and download Hugging Face models, and manage them with live progress.
- Unified provider authentication: Centralized model configuration and provider login across Pi, with file-backed dynamic catalogs and Radius gateway support.
- Built-in provider extensions: Extensions can now register complete Pi providers with authentication, model refresh, filtering, and custom streaming.
- Simplified sign-in: OpenRouter and Kimi Code now support /login for authorization without manual API key entry; headless SSH login also available.
Workflow and Credentials
- Message copy shortcut: Press Ctrl+X to copy the last assistant message or any selected message in the conversation tree.
- Credential export: Export API keys and bearer tokens for external clients with automatic OAuth refresh and configurable token validity.