Anthropic - PandaProbe

Installation

pip
uv

pip install "pandaprobe[anthropic]"

uv add "pandaprobe[anthropic]"

Setup

Sync
Async

from pandaprobe.wrappers import wrap_anthropic
from anthropic import Anthropic

client = wrap_anthropic(Anthropic())

from pandaprobe.wrappers import wrap_anthropic
from anthropic import AsyncAnthropic

async_client = wrap_anthropic(AsyncAnthropic())

Messages API

Span name: "anthropic-messages", SpanKind: LLM

response = client.messages.create(
    model="claude-sonnet-4-6",
    max_tokens=1024,
    system="You are a helpful assistant.",
    messages=[{"role": "user", "content": "Explain quantum computing."}],
)

What gets traced

Input: system prompt (from top-level system kwarg) plus messages, normalized to a standard format
Output: text content blocks
Model name
Token usage
Model parameters: temperature, top_p, top_k, max_tokens, stop_sequences, thinking configuration
Extended thinking or reasoning blocks stored in metadata as reasoning_summary

Streaming

response = client.messages.create(
    model="claude-sonnet-4-6",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Hello!"}],
    stream=True,
)
for event in response:
    if hasattr(event, "delta") and hasattr(event.delta, "text"):
        print(event.delta.text, end="")

with client.messages.stream(
    model="claude-sonnet-4-6",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Hello!"}],
) as stream:
    for text in stream.text_stream:
        print(text, end="")

Both streaming patterns are fully supported with time-to-first-token tracking.

Extended thinking

When using Anthropic’s extended thinking feature, thinking blocks are automatically extracted and stored in the span metadata under the reasoning_summary key. Thinking content is stripped from the visible output.

Token usage mapping

Anthropic Field	PandaProbe Field
`input_tokens`	`prompt_tokens`
`output_tokens`	`completion_tokens`
`cache_read_input_tokens`	`cache_read_tokens`
`cache_creation_input_tokens`	`cache_creation_tokens`