Gpt-5.4-mini: 66K prompt tokens for a 1920x1080 PNG

API
, ,
1

I’m sending a properly formatted image_url request to gpt-5.4-mini via the Chat Completions API. The image is a 1920x1080 PNG (~131KB), sent as a base64 data URL with detail: high.

Per the Images and vision docs, gpt-5.4-mini uses patch-based image tokenization with a 1,536-patch budget and a 1.62x multiplier. A 1920x1080 image should cost approximately 2,400 prompt tokens.

Instead, I’m seeing ~66,000 prompt tokens. This matches almost exactly what you’d get if the base64 string were tokenized as text (~131KB PNG → ~175KB base64 → ~66K text tokens at ~4 chars/token).

Also, the prompt_tokens_details in the API response contains no image_tokens field:

"prompt_tokens_details": {
    "audio_tokens": 0,
    "cached_tokens": 2304
}

Shouldn’t this show image_tokens if I’m sending an image?

Request payload (base64 truncated):

{
  "model": "gpt-5.4-mini",
  "max_completion_tokens": 256,
  "messages": [
    {
      "role": "user",
      "content": [
        {
          "type": "text",
          "text": "What do you see? Reply in one sentence."
        },
        {
          "type": "image_url",
          "image_url": {
            "url": "data:image/png;base64,iVBORw...",
            "detail": "high"
          }
        }
      ]
    }
  ]
}

The model does understand the image. It returns a correct description of the screenshot contents. So the vision capability works, but tokenization/billing appears to fall back to treating the base64 as plain text.

Is this a known issue with gpt-5.4-mini on the Chat Completions API?

Here’s a minimal script that reproduces this, hitting the API directly (no SDK):

#!/usr/bin/env bash
set -euo pipefail

TMPFILE=$(mktemp)
trap 'rm -f "$TMPFILE"' EXIT

BASE64_IMAGE=$(base64 -w 0 "<path_to_1080p_image>")

cat > "$TMPFILE" <<EOF
{
  "model": "gpt-5.4-mini",
  "max_completion_tokens": 256,
  "messages": [
    {
      "role": "user",
      "content": [
        {"type": "text", "text": "What do you see? Reply in one sentence."},
        {
          "type": "image_url",
          "image_url": {
            "url": "data:image/png;base64,${BASE64_IMAGE}",
            "detail": "high"
          }
        }
      ]
    }
  ]
}
EOF

curl -s -w "\nHTTP_STATUS: %{http_code}\n" \
  https://api.openai.com/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer ${OPENAI_KEY}" \
  -d @"$TMPFILE"

And the response:

thiagolobo@nephtis-desktop:~/$ ./openai.sh 
{
  "id": "chatcmpl-DdPgVQgzShKGIbQoqun4PyrV6zywL",
  "object": "chat.completion",
  "created": 1778285895,
  "model": "gpt-5.4-mini-2026-03-17",
  "choices": [
    {
      "index": 0,
      "message": {
        "role": "assistant",
        "content": "A browser is open to the Van Zandt CAD online property tax search page with search fields and helpful hints visible.",
        "refusal": null,
        "annotations": []
      },
      "finish_reason": "stop"
    }
  ],
  "usage": {
    "prompt_tokens": 65599,
    "completion_tokens": 26,
    "total_tokens": 65625,
    "prompt_tokens_details": {
      "cached_tokens": 1792,
      "audio_tokens": 0
    },
    "completion_tokens_details": {
      "reasoning_tokens": 0,
      "audio_tokens": 0,
      "accepted_prediction_tokens": 0,
      "rejected_prediction_tokens": 0
    }
  },
  "service_tier": "default",
  "system_fingerprint": null
}

HTTP_STATUS: 200
2

We’re seeing similarly strange behavior.

I benchmarked the same prompt (short question, single 1920x1080 jpg) across a handful of models and got wildly different input token counts:

image
image454×115 2.73 KB

3

Welp. Whatever was happening seems to now be fixed. I’m now getting this response after running the very same script:

thiagolobo@nephtis-desktop:~/Documents/pythonicai/tools/lib/ds/title/taxcard$ ./openai.sh 
{
  "id": "chatcmpl-DdlkEzQZ9f68WN38pTYAcJuQ97euO",
  "object": "chat.completion",
  "created": 1778370694,
  "model": "gpt-5.4-mini-2026-03-17",
  "choices": [
    {
      "index": 0,
      "message": {
        "role": "assistant",
        "content": "A browser is open to the Van Zandt CAD online property tax search page with search filters and helpful hints visible.",
        "refusal": null,
        "annotations": []
      },
      "finish_reason": "stop"
    }
  ],
  "usage": {
    "prompt_tokens": 2358,
    "completion_tokens": 26,
    "total_tokens": 2384,
    "prompt_tokens_details": {
      "cached_tokens": 1792,
      "audio_tokens": 0
    },
    "completion_tokens_details": {
      "reasoning_tokens": 0,
      "audio_tokens": 0,
      "accepted_prediction_tokens": 0,
      "rejected_prediction_tokens": 0
    }
  },
  "service_tier": "default",
  "system_fingerprint": null
}

HTTP_STATUS: 200