BinaryImage causes token explosion when passed to marvin.cast()

[zzstoatzz]

Problem

When passing a BinaryImage to marvin.cast(), the resulting API request exceeds token limits even for moderately-sized images.

Reproduction

Using a 1600x1200 PNG image (~200KB on disk):

from pathlib import Path
from pydantic_ai.messages import BinaryImage
import marvin
from pydantic import BaseModel

class ModerationResult(BaseModel):
    is_safe: bool
    explanation: str

image_bytes = Path("handguns.png").read_bytes()
binary_image = BinaryImage(data=image_bytes, media_type="image/png")

moderator_agent = marvin.Agent(
    name="Content Moderator",
    model="anthropic:claude-sonnet-4-5",
    model_settings={"temperature": 0},
)

result = marvin.cast(
    binary_image,
    target=ModerationResult,
    instructions="Analyze this image",
    agent=moderator_agent,
)

Error:

anthropic.BadRequestError: Error code: 400 - {'type': 'error', 'error': {'type': 'invalid_request_error', 'message': 'prompt is too long: 207768 tokens > 200000 maximum'}}

Expected Behavior

The same image works perfectly when using pydantic-ai directly:

from pydantic_ai import Agent
from pydantic_ai.messages import BinaryImage

moderator_agent = Agent[None, ModerationResult](
    "gateway/anthropic:claude-sonnet-4-5",
    output_type=ModerationResult
)

result = await moderator_agent.run(
    ["Analyze this image", binary_image],
    instructions="Analyze this image",
    model_settings={"temperature": 0},
)
# Works fine - no token issues

Analysis

Something in marvin's request construction when handling BinaryImage is causing the image data to be sent in a way that consumes 207K+ tokens, when the same image via pydantic-ai directly uses a normal amount of tokens.

Possibly marvin is encoding the image differently, duplicating it in the request, or not properly utilizing Anthropic's image handling?

Environment

• marvin: 2.3.9
• pydantic-ai: latest
• anthropic: latest (via pydantic-ai)
• model: claude-sonnet-4-5

Impact

This makes image analysis completely unusable via marvin.cast() for any real-world images, as even small images exceed token limits.