Streaming with base64 images inflates input token count

[bogoconic1]

Bug Description

When using client.messages.stream() with base64 images in message content, the OpenTelemetry span reports inflated input_tokens. Non-streaming client.messages.create() reports correctly.

Reproduction

import base64, io, os
from anthropic import Anthropic
from opentelemetry.instrumentation.anthropic import AnthropicInstrumentor
from PIL import Image

AnthropicInstrumentor().instrument()

client = Anthropic(api_key=os.environ["ANTHROPIC_API_KEY"])

img = Image.new("RGB", (500, 500), color=(200, 200, 255))
buf = io.BytesIO()
img.save(buf, format="PNG")
b64 = base64.b64encode(buf.getvalue()).decode()

messages = [{
    "role": "user",
    "content": [
        {"type": "image", "source": {"type": "base64", "media_type": "image/png", "data": b64}},
        {"type": "text", "text": "What color is this? One word."},
    ],
}]

# Non-streaming
response = client.messages.create(model="claude-haiku-4-5-20251001", max_tokens=100, messages=messages)
print(f"[CREATE] API usage: input={response.usage.input_tokens}")

# Streaming
with client.messages.stream(model="claude-haiku-4-5-20251001", max_tokens=100, messages=messages) as stream:
    response = stream.get_final_message()
print(f"[STREAM] API usage: input={response.usage.input_tokens}")

Observed Behavior

Method	API usage.input_tokens	OTel span reported
create()	343	343
stream()	343	1,633

Expected Behavior

OTel span should match the API's usage.input_tokens for both methods.

Environment

• opentelemetry-instrumentation-anthropic: 0.57.0
• anthropic: 0.89.0
• Python 3.11

[carsonjc04]

Hi, I'd like to work on this issue if available! From what I found the issue lies in how OpenLLMetry reconstructs usage from the SSE events. The streaming code only sets token attributes when enrich_token_usage is True. When it's False (default), no gen_ai.usage.input_tokens attribute is set on the span at all, even though the API did provide the real number in the stream events.

This causes downstream impact on Langfuse where it sees the span without gen_ai.usage.input_tokens so they fall back to estimating tokens using a tokeniser on the raw content which includes the base 64 image string that'll drastically increase an incorrect token value.

Fix: We should always read the token counts from the stream events and set them on span, regardless of the enrich_token_usage flag. The enrich_token_usage flag will still work properly for non-streaming, but this bug is causing it to rely on the estimation tokeniser instead of the given api usage.