Adding image, audio, or video inputs to an LLM application introduces a different cost structure than text-only usage, since providers typically convert non-text inputs into an equivalent token count using their own conversion methodology rather than billing them as a flat per-item fee. Understanding that conversion is essential before estimating cost for a multimodal feature.
How Image Inputs Get Converted to Tokens
Providers typically translate an image into a token count based on its resolution and, in some cases, how the image is tiled or segmented internally for processing, meaning a larger or higher-resolution image consumes meaningfully more tokens than a smaller one, similar in spirit to how a longer text passage consumes more tokens than a shorter one.
This makes image resolution a direct, controllable cost lever: resizing or compressing an image to the minimum resolution actually needed for the task, rather than sending a full, high-resolution original by default, is one of the most straightforward multimodal cost optimizations available and often has no meaningful effect on task accuracy.
Document and Visual QA Workloads
Document processing use cases, such as extracting data from scanned forms, invoices, or receipts, often involve sending an entire page image even when only a small region contains the relevant information. Where feasible, cropping to the relevant region or splitting a multi-page document into targeted sections before sending it can reduce token cost substantially compared to sending full, unmodified pages.
For high-volume document pipelines, evaluate whether a specialized OCR or document-extraction service, potentially cheaper and more accurate for pure text extraction than a general-purpose multimodal LLM call, should handle the extraction step, reserving multimodal LLM calls for tasks that genuinely require visual reasoning beyond text extraction alone.
Multiple Images in a Single Request
Sending several images in one request, such as for comparison tasks or multi-page document analysis, multiplies the per-image token cost across every image included, so the same resolution-reduction and relevance-filtering principles apply with even greater impact as image count per request grows.
Consider whether a task genuinely needs multiple images simultaneously in context or whether it can be decomposed into sequential single-image calls with intermediate results combined afterward, since the right decomposition depends on whether the task requires cross-image reasoning or can be handled independently per image.
Audio and Video Considerations
Audio and video inputs are typically converted to tokens based on duration rather than resolution, meaning cost scales with the length of the clip processed. Trimming input to only the relevant time range, rather than submitting an entire long recording when only a portion is relevant, is the direct analog of image cropping for these modalities.
For audio-heavy workloads such as transcription or summarization of long recordings, evaluate whether a dedicated speech-to-text service, followed by a text-based LLM call on the transcript, is more cost-effective than sending raw audio directly to a multimodal model, since text tokens are typically far cheaper than the audio-token equivalent of the same content.
Key takeaways
- Resize or compress images to the minimum resolution actually needed before sending them to the API.
- Crop document images to the relevant region instead of sending full, unmodified pages by default.
- Consider a dedicated OCR or speech-to-text service ahead of the LLM call for pure extraction tasks.
- Account for token cost multiplying across every image when sending multiple images in one request.
- Trim audio and video inputs to the relevant time range rather than submitting full-length recordings.
Bottom line
Multimodal inputs introduce a cost dimension, resolution and duration, that has no direct analog in text-only workloads, and controlling it is often as simple as trimming inputs to what the task actually requires. Teams building image- or audio-heavy features should model this conversion explicitly rather than assuming multimodal costs behave like text costs.