Aider v0.76.0: Reasoning Models & Notifications

TL;DR

• Full support for thinking/reasoning models with token budget control via --thinking-tokens
• Desktop notifications when LLM responses are ready with --notifications flag
• Added Claude 3.7 Sonnet, GPT-4.5-preview, QWQ 32B, and DeepSeek V3 free tier support
• Deprecated model shortcut args like --4o and --opus in favor of --model

New

• Thinking/reasoning model support — Control token budget with --thinking-tokens, display reasoning content from LLMs, and improved handling of reasoning tags.
• Desktop notifications — Get notified when LLM responses are ready with --notifications flag; customize notification command with --notifications-command.
• Claude 3.7 Sonnet — Now available on OpenRouter (default), Bedrock, and Vertex AI, including beta variant.
• GPT-4.5-preview — Added support for OpenAI's latest preview model.
• QWQ 32B and DeepSeek V3 free — QWQ 32B support added; DeepSeek V3 available free on OpenRouter via --model openrouter/deepseek/deepseek-chat:free.
• Tree-sitter language pack — Switched to tree-sitter-language-pack for improved language support.

Fixed

• Git error handling when reading staged files — Now properly handles Git errors without crashing.
• Git identity retrieval — Respects global Git configuration settings.
• EOF handling (Ctrl+D) — Improved error handling in user input prompts.
• Empty LLM responses — Clearer warning messages when models return empty responses.
• SSL verification — Better control for model information requests.
• Weak model name patterns — Fixed patterns to match main model names for consistency.

Deprecated

• Model shortcut args — --4o, --opus, and similar shortcuts are deprecated; use --model flag instead.
• remove_reasoning setting — Replaced by reasoning_tag; deprecation warning added.

Update with: pip install --upgrade aider-ai

Source: Aider