🐛 Bug Report: Anthropic: since anthropic 1.14.0 each call reports its usage on two spans

オープン
#4,700 コメント 4 件 リアクション 0 件 担当者 0 名 GitHub で見る

メンテナーはふだん 1 日以内に返信

@simin-yuan がすでに取り組んでいます。

2026年11月12日 から。

  • #4703 @simin-yuan による — オープン

評価

難易度
3/5
見積もり時間
1〜2日
初心者へのやさしさ
56/100
issue の種類
バグ
明瞭さ
おおむね明確
活発さ
活発
技術スタック
python

調査の方向性

Start by reading the Anthropic instrumentation entry point where instrument() configures SDK behavior, then check the tests for environment variable handling and client creation order. Confirm the existing SDK switch prevents duplicate usage spans while preserving a caller's explicit setting, and verify the documented cases around clients created before instrumentation and streaming failures.

索引モデルが issue の本文から書いたものです。

説明

Which component is this bug for?

Anthropic Instrumentation

📜 Description

Since anthropic 1.14.0 the SDK starts its own span for each call by default when a global tracer provider is registered. With this instrumentation on top, each call reports its token usage on two spans, anthropic.messages.create and anthropic.chat, in two separate traces. A call with 1,000 input and 200 output tokens then costs $0.012 in the project totals of Phoenix 21.0.0 and Opik 2.2.143, where it should be $0.006 (raw responses, r2-A-lllatest, o1-A-lllatest).

👟 Reproduction steps
# double_span.py
import json
import threading
from http.server import BaseHTTPRequestHandler, HTTPServer

import anthropic
from opentelemetry import trace
from opentelemetry.instrumentation.anthropic import AnthropicInstrumentor
from opentelemetry.sdk.trace import TracerProvider
from opentelemetry.sdk.trace.export import SimpleSpanProcessor
from opentelemetry.sdk.trace.export.in_memory_span_exporter import InMemorySpanExporter


class FakeAnthropic(BaseHTTPRequestHandler):
    def do_POST(self):
        self.rfile.read(int(self.headers["content-length"]))
        body = json.dumps({
            "id": "msg_1", "type": "message", "role": "assistant", "model": "claude-sonnet-5-5",
            "content": [{"type": "text", "text": "hi"}],
            "stop_reason": "end_turn", "stop_sequence": None,
            "usage": {"input_tokens": 1000, "output_tokens": 200},
        }).encode()
        self.send_response(200)
        self.send_header("content-type", "application/json")
        self.send_header("content-length", str(len(body)))
        self.end_headers()
        self.wfile.write(body)

    def log_message(self, *args):
        pass


server = HTTPServer(("127.0.0.1", 0), FakeAnthropic)
threading.Thread(target=server.serve_forever, daemon=True).start()

exporter = InMemorySpanExporter()
provider = TracerProvider()
provider.add_span_processor(SimpleSpanProcessor(exporter))
trace.set_tracer_provider(provider)
AnthropicInstrumentor().instrument()

client = anthropic.Anthropic(api_key="x", base_url=f"http://127.0.0.1:{server.server_port}")
client.messages.create(
    model="claude-sonnet-5-5", max_tokens=10, messages=[{"role": "user", "content": "hi"}]
)

for span in exporter.get_finished_spans():
    tokens = {k: v for k, v in span.attributes.items() if k in ("gen_ai.usage.input_tokens", "gen_ai.usage.output_tokens")}
    print(span.name, "|", span.instrumentation_scope.name, "|", tokens)
# anthropic 2.6.0, opentelemetry-instrumentation-anthropic 0.63.0, opentelemetry-sdk 1.45.2, Python 3.12
$ python double_span.py
anthropic.messages.create | com.anthropic.sdk.python | {'gen_ai.usage.input_tokens': 1000, 'gen_ai.usage.output_tokens': 200}
anthropic.chat | opentelemetry.instrumentation.anthropic | {'gen_ai.usage.input_tokens': 1000, 'gen_ai.usage.output_tokens': 200}
$ ANTHROPIC_OPEN_TELEMETRY=false python double_span.py
anthropic.chat | opentelemetry.instrumentation.anthropic | {'gen_ai.usage.input_tokens': 1000, 'gen_ai.usage.output_tokens': 200}
👍 Expected behavior

One span per call reports usage. The fix I'd pick: instrument() calls os.environ.setdefault("ANTHROPIC_OPEN_TELEMETRY", "false"). It uses the SDK's public switch, and a caller's own setting wins. Added after instrument() in the script above, it gives the anthropic.chat span alone. It does not reach a client created before instrument().

👎 Actual Behavior with Screenshots

Two spans with usage, as above.

🤖 Python Version

3.12

📃 Provide any additional context for the Bug.
  • Until #4536 and #4643 land, the SDK's span is the only record of a stream the app leaves early or of a sync stream that fails, so those should land first.
  • Arize-ai/openinference#4121 reaches every client by wrapping a private SDK function. Here it needs anthropic.chat current first, as #4366 does.
👀 Have you spent some time to check if this bug has been raised before?
  • I checked and didn't find similar issue
Are you willing to submit PR?

Yes I am willing to submit a PR!

主要言語
Python
スター
7.5k
フォーク
1.2k
平均マージ
1日 23時間
マージ済み PR(30日)
3

環境構築

はじめの一歩

  1. issue を最後まで読み、次にプロジェクトのコントリビューションガイドを読みます。
  2. 着手することを issue にコメントします — 二人が同じ作業をするのを防げます。
  3. リポジトリをフォークし、ブランチを切って変更します。
  4. issue 番号を参照したプルリクエストを送ります。

traceloop/openllmetry のほかの issue

traceloop/openllmetry の issue をすべて見る

似ている issue

Python の issue をもっと見る

新しい issue をメールで受け取る

初心者向けの GitHub issue を短くまとめたダイジェスト。