How AI counts pictures, speech and video

A picture has no useful word count until somebody describes it. A recording contains speech, pauses and sounds beyond its transcript. A video adds a changing sequence of images. AI services therefore convert different media into the representations their models use, then apply the charging rules for the chosen operation.

The examples below concern developer APIs, using documentation checked on 12 October 2026. We calculated the arithmetic locally; we did not send these examples to a paid endpoint. The figures isolate particular components, so they are not estimates of a complete project or a consumer subscription allowance.

A picture becomes a sized input

OpenAI’s gpt-6-astra covers processed images with 32-by-32-pixel patches. Under high detail, its 2,500-patch budget can trigger resizing. Low detail first fits the picture within 512 by 512 pixels. A 1.2 multiplier, rounded up, converts the resulting patch count into billable image input tokens.[1]

The difference comes from the representation supplied to the model. Compressing a JPEG into fewer file bytes while retaining the same dimensions would not change this patch arithmetic. Reducing its dimensions does. Neither calculation establishes whether the smaller version preserves the detail needed for your task.[1]

For a photograph of a room, the broad layout may be enough. For a photograph of a timetable, the small lettering may be the whole point. Decide what the answer must contain before reducing the material available to produce it.

Speech has separate input and output units

OpenAI’s Realtime documentation assigns user audio one token per 100 milliseconds and assistant audio one per 50 milliseconds. The same speaking duration therefore contributes different token counts depending on direction. Its current getting-started example uses gpt-realtime-2.1.[2][4]

For that model, the standard audio rates are US$32 per million input tokens and $64 per million output tokens. The following calculation assumes 45 seconds of user audio accepted as input and a 15-second spoken answer, each counted once.[3]

These are speaking durations, not the wall-clock length of a call. Voice activity detection can exclude empty input audio. Conversation history can also be charged again as input on later responses, with caching affecting eligible portions. Check the response.done usage details when assessing an actual session.[2]

Even another OpenAI voice product uses a different basis: GPT-Live 1 charges for session duration, with backend model and tool work separately billed. Its listed voice rate is $0.05 per minute. Selecting “voice AI” is therefore not enough to identify the unit you will buy.[3]

For video, ask what was loaded

Gemini’s default static video processing samples one frame per second and includes audio and timestamps. Gemini 3.8 Flash also supports agentic processing: the model can navigate a recording and load selected frames, audio or transcript as required. Frame resolution and the material selected affect usage.[5]

A two-minute recording sampled at one frame per second supplies about 120 sampled frames, before considering its soundtrack and metadata. That is our duration calculation, not a token count. Counting every frame stored inside the video file would describe a different input. With agentic processing, the requested question also influences which material is inspected.[5]

This is why “What happens in this clip?” and “Read the serial number visible near the end” can justify different processing choices. A useful product should connect the answer to the relevant moment, so the person checking it can inspect the same evidence.

Generating a video buys a different operation

Veo 3.1 generates eight-second videos with audio through the Gemini API. Google’s paid rate for the Standard variant at 1080p is $0.40 per generated second. One successful eight-second generation therefore costs $3.20 at that rate; three successfully generated alternatives cost $9.60, even if you keep only one.[6][7]

Those are output-generation charges. They do not price analysing an existing eight-second clip. Google states that Veo is charged only when a video is successfully generated; a clip you receive but dislike is still a generated result.[7]

Read the usage record for the chosen job

Google documents counting inputs before a request and inspecting usage after generation. Keep the exact model identifier, input dimensions or duration, processing settings and output requirement beside those figures. That makes an estimate reproducible and lets you investigate a change in the bill.[8]

For an original trial, use three examples from your own work: a product photograph, a spoken enquiry and a short demonstration. Specify the answer each must deliver, then compare settings on that task. Record both the cost and whether the result preserved the useful detail. A smaller count is valuable when the completed work still does what you need.

Report an error or suggest a correction