When a screenshot, a voice memo or an hour of meeting video reaches an API invoice, nobody charges by pixel, second or frame. The provider first converts the media into ordinary input tokens, then applies the per-million rate printed on its pricing page. The three major labs run three different converters, and the conversion step moves the cost of the same file by far more than the rate does.
Three rulebooks
OpenAI tokenizes images as 32x32 pixel patches on its newest models. For gpt-6-astra at high detail, the count is the number of patches times a 1.2 multiplier, with a 2,500-patch budget and a 65,535-pixel maximum dimension. The company's own examples: a 1024x1024 image costs 1,229 tokens, a 2048x2048 image is squeezed to 1600x1600 and costs 3,000 tokens, and a 4096x512 panorama costs 2,458 tokens. Older models still on tile pricing are cheaper per image: gpt-5.1 charges 70 base tokens plus 140 per 512-pixel tile, so a typical screenshot lands in the hundreds of tokens. Hard limits apply per file: any image that needs more than 30,000 patches after resizing is rejected rather than shrunk, and a request can carry up to 1,500 images in a 512 MB payload.
Anthropic charges by visual token, defined as one 28x28 pixel patch, so an image costs ceil(width/28) x ceil(height/28) at native resolution. What differs is the ceiling. On standard-tier models the image is downscaled to at most 1,568 pixels on the long edge and 1,568 visual tokens. Claude 4.7 and later add a high-resolution tier that allows 2,576 pixels and 4,784 visual tokens, roughly three times the standard tier's bill for the same picture. Anthropic's worked example: a 1000x1000 image costs 1,296 tokens on either tier, but a 4K frame costs 1,560 tokens on a standard-tier model and 4,784 on a high-resolution one.
Google stops deriving tokens from resolution altogether and sells the dial. On Gemini 3 and 4 models, each image gets a fixed token budget set by its media_resolution level: 280 tokens at low, 560 at medium, 1,120 at high, with 1,120 as the default. The image is downscaled to fit whichever budget you pick, so a 12-megapixel photo and a 1-megapixel photo cost the same by default. A 2,240-token ultra_high tier exists and is the required setting for computer-use workloads. PDF pages follow the same dial (280, 560 or 1,120) but plus the page's native text tokens, so digital PDFs are cheap to ingest.

Where the meters agree
For the reference case of one megapixel, the three rulebooks converge within about 15 percent. A 1000x1000 image costs 1,229 tokens on gpt-6-astra, 1,296 on Claude, and 1,120 on Gemini's default tier. Three labs, three formulas, one answer: a picture is worth roughly 1,200 tokens.
The dollar column is where providers separate, and it is mostly the model rate talking, not the meter. At standard rates:
- gpt-6-astra input is $10 per million tokens, so a 1-megapixel image runs about $12.29 per thousand images.
- Claude Opus 5.5 input is $4 per million; Anthropic itself quotes $5.18 per thousand 1-megapixel images, and $19.14 per thousand 4K images on its high-resolution tier. Haiku 4.5 at $1 per million brings the same 1-megapixel image down to $1.30 per thousand.
- Gemini 4 Hydrogen input is $1.50 per million, about $1.68 per thousand images at the default tier; Gemini 3.5 Flash-Lite at $0.30 per million drops it to $0.34 per thousand, and the low tier to $0.08.
Chosen well, the same 1-megapixel photo spans $0.08 per thousand images (Gemini low tier on a Flash-Lite model) to $12.29 per thousand (gpt-6-astra, high detail). The per-million rate is the wide lever across models, the resolution dial is a 4x lever inside one model, and pre-resizing decides whether you pay at all for pixels the model discards.
Audio is a per-second meter

Audio does not scale with resolution at all, but the three vendors price it very differently. Gemini tokenizes audio at a fixed 25 tokens per second on every model and every resolution level, which works out to 1,500 tokens per minute; Google's own pricing footnotes land file transcription at about $0.005 per minute blended, live transcription near $0.009, and Live API conversations near $0.0368. OpenAI's realtime tier prices audio as its own modality: on gpt-realtime-2.1 audio input costs $32 per million tokens against $4 for text, an 8x premium, while cached audio input drops to $0.40, a 98.75 percent discount. Its GPT-Live-1 model sidesteps token math entirely at $0.05 per minute, billed by the second, with backend usage charged separately.
Video is the cheapest multimodal input per minute if you let the provider sample frames. Gemini takes one frame per second by default at 70 tokens each, so a one-hour video is roughly 252,000 video tokens plus 90,000 audio tokens, about $0.51 at Hydrogen rates. Raising the same clip to high resolution for on-screen text pushes frames from 70 to 280 tokens, quadrupling the visual bill. Custom frame rates and an agentic processing mode exist precisely because 1 fps of full frames is often wasted money.
The computer-use tax
Agent loops are where input billing gets expensive, because screenshots pass the meter on every step. A 1920x1080 screenshot costs about 2,448 tokens on gpt-6-astra at high detail, 2,691 on a Claude high-resolution-tier model, and 2,240 on Gemini's required ultra_high setting. All three labs steer you away from casual resizing in this loop: Anthropic rejects oversized screenshots returned through tool_result instead of downscaling them, and counts resent history images toward its many-image limits. OpenAI recommends original detail for coordinate-sensitive work. Meanwhile the conversation history re-sends every earlier screenshot, which is why the cache discounts matter more than the base rate: a 90 to 98 percent cached-input discount turns a resending loop from expensive into tolerable.
What to do about the meter
- Resize before upload, not after. Anthropic recommends it explicitly; OpenAI rejects images above 30,000 patches instead of shrinking them.
- Pick the resolution tier per image, not per account. Google supports per-content-item settings on Gemini 3 and 4; mixing high for one diagram and low for context images is the intended pattern.
- Cache media-heavy prefixes. Cached rates on all three platforms are 90 percent or more below standard.
- Budget by tokens, not requests. Image tokens count toward OpenAI's tokens-per-minute limits, so a batch of 1,500 screenshots can trip a rate limit that text would not.
The pricing pages will keep changing the per-million numbers. The 1,200-token picture is the stable part of the bill.
