Image analysis

Models with image support accept graphics through the content array with a type: "image_url" element. The image address is either a data: URI with base64 data or an https link. In the GET /v1/models catalog such models carry the capabilities.vision: true flag.

Images are processed for free by bitrix/bitrixgpt-5.5 and bitrix/bitrixgpt-5.5-thinking. With your own provider key (BYOK), Anthropic Claude models are available: anthropic/claude-opus-4-6, anthropic/claude-sonnet-4-6, anthropic/claude-haiku-4-5-20251001. For these models, the platform automatically converts the image_url element to the Anthropic format — nothing needs to change on the client side.

Request example

curl -X POST https://vibecode.bitrix24.com/v1/chat/completions \
  -H "X-Api-Key: YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "bitrix/bitrixgpt-5.5",
    "messages": [
      {
        "role": "user",
        "content": [
          {"type": "text", "text": "Describe what is in the image"},
          {"type": "image_url", "image_url": {"url": "data:image/jpeg;base64,/9j/4AAQSkZJRg..."}}
        ]
      }
    ],
    "max_tokens": 500
  }'

Limits

MiB is a mebibyte, a binary unit per the IEC 80000-13 standard: 1 MiB = 2²⁰ = 1 048 576 bytes ≈ 1.05 MB. Limits are checked as N × 1024 × 1024, so the error message also states the size in MiB — for example, decoded size 21.3 MiB exceeds limit 20 MiB.

Content check

Before the request is sent to the model, the content of every image_url element passed as a data: URI is inspected by its actual bytes rather than its declared MIME type. If the content is not an image — for example an HTML page or an error response encoded in base64 and declared as image/png — the element is not dropped but replaced in place with the text placeholder [image unavailable: <reason>]. The remaining images are processed normally and the request succeeds.

The content check applies to data: URIs only. A candidate passed as an https:// link is forwarded to the model as is, without fetching or inspecting its bytes — validate such links on your side.

Positions are preserved: the length of the content array does not change, so any candidate numbering on your side stays correct.

The substitution is visible in the response:

• the X-Image-Parts-Rejected header carries the number of replaced elements — this is the single signal available in every response mode; for streaming responses it arrives with the start of the stream;
• in a regular (non-streaming) response the warnings field additionally carries an entry with the code IMAGE_CONTENT_REJECTED and the substitution details. A streaming response has no warnings field — rely on the header.

Errors

Content errors (not an image, corrupted base64, unsupported MIME type, over 20 MiB) no longer fail the request — such an element is replaced with a placeholder (see "Content check").

For the full list of common API errors, see Errors.

Known specifics

The content array is only needed for images. For text requests pass content as a string. An array with a single text element also works but is redundant.

For bitrix/* models, the platform downscales large images itself. When the request body contains images and exceeds the limit, the platform downscales its data: images, starting with the largest, to 1536 pixels on the long side and saves them as JPEG at quality 80. If that is not enough, it downscales them to 1024 pixels at quality 70. An image with transparent pixels is downscaled as PNG with its transparency kept, and only if that does not fit the limit is it saved as JPEG on a white background. A GIF keeps only its first frame; images passed as an https link are left unchanged. A body that is still over the limit gets a 413. With stream: true, the ai_payload_too_large and ai_image_resize_busy refusals arrive as an error frame inside a 200 response, and the retry delay comes in the frame's retryAfter field rather than in a header. An image over 20 MiB after decoding is not downscaled: it is replaced with a placeholder, as described in "Content check".

To control the quality, downscale images yourself. A 12 MP JPEG photo takes 4–5 MB. Downscaled to 1500 pixels on the long side and saved as JPEG or WebP at quality 80–85, it takes 0.2–1 MB and reaches the model unchanged.

Image size is checked after decoding. The 20 MiB limit applies to the image itself, not to the base64 string. The string is roughly a third longer, so the request body for models on your own provider key is limited to 30 MiB.

See also

• Create a chat completion
• List of models
• Your own keys (BYOK)