Python SDK 1.45.0

Features

• Aligned offline/SDK evaluations with tenant-configured evaluation types including rule-engine context resolution and per-type thresholding (#1382)
• Added OpenLIT instrumentation for OpenAI Responses.parse and AsyncResponses.parse methods using existing response wrappers (#1383)
• Added cached token pricing support across multiple SDK instrumentations including AI21, Azure AI Inference, Bedrock, Cohere, Google AI Studio, Groq, Mistral, OpenAI, PremAI, Pydo, Reka, Sarvam, Strands, and Vertex AI (#1390)

Fixes

• Fixed OpenTelemetry context detach errors in Anthropic streaming by adding safe_detach helper that checks task context before detaching (#1370)
• Fixed cost calculation for cached tokens in litellm and claude_agent_sdk instrumentations to use cache rates instead of full prompt price (#1379)
• Fixed AssemblyAI transcription cost calculation to use audio duration instead of URL length (#1381)
• Fixed Anthropic instrumentation to correctly capture all parallel tool_use calls in both streaming and non-streaming responses (#1399)
• Fixed Gemini google_ai_studio instrumentation to correctly capture multiple parallel function_call parts in both streaming and non-streaming responses (#1401)

Dependencies

• Updated langchain-core requirement to allow version 1.5.0 and above (#1363)
• Updated crewai requirement to allow version 1.15.5 and above for Python 3.10+ (#1388)
• Updated langchain requirement to allow version 1.3.14 and above while staying below 2.0.0 for Python 3.10+ (#1389)

Documentation

• Added mem0 agent memory tracing example with runnable demo and improved mem0 instrumentation documentation (#1374)