Word by word, a spoken description of your screen costs an AI coding agent roughly 150 tokens. A single screenshot costs about 2,691 on Claude 4.7 and later. Frame-dumping 30 seconds of video into stills costs 80,730 tokens, about 40% of a 200k context window gone before the agent even reads your bug report. Words are cheapest by a wide margin, but only when they name an exact location.
Updated August 2026. This breakdown uses Anthropic's own image-token formula plus measured figures from a real test bundle, not estimates. It's part of the wider picture of Claude Code token costs: once you see where the tokens actually go, video, screenshots and words turn out to have very different price tags for describing the same bug, and the cheapest one isn't automatically the best one.
- 80,730 tokens — to frame-dump a 30-second review at 1080p, 1 fps
- 2,691 tokens — for one 1080p screenshot (Claude 4.7+)
- ~150 tokens — for a typical spoken review, transcribed
Can an AI agent watch a video?
No. Claude Code's context accepts text and images, not video files. The only way to show an agent a screen recording is to convert it into a sequence of still frames, each billed as an image at Anthropic's per-frame token rate. "Video vs. text" is really a question of how many image-frames you're paying for versus how much prose.
Nobody actually pastes a raw .mp4 into an agent's context window; there's no code path for it. What happens in practice is frame extraction: a tool samples the recording at some interval, usually once per second, and hands the agent that stack of stills instead. So "how much does video cost" collapses into "how many frames, at what resolution" — which is a question the next section can answer exactly. For the full mechanics of why that conversion step is unavoidable, see why your coding agent can't watch a screen recording.
What does a frame-dump cost?
A 30-second screen review captured at one frame per second is 30 still images. At the 1080p rate of 2,691 tokens per frame on Claude 4.7 and later, that's 80,730 tokens, or about $0.81 at $10 per million input tokens. Nothing in that total tells the agent which of the 30 frames actually matters.
The per-frame number comes straight from Anthropic's published patch formula: an image costs ceil(width / 28) × ceil(height / 28) visual tokens (Source: Anthropic, 2026). For a full 1920×1080 frame, that's ceil(1920/28) × ceil(1080/28) = 69 × 39 = 2,691 tokens on Claude 4.7 and later, exact, no rounding. Multiply by 30 frames for one second-by-second minute of footage and the total is 80,730 tokens, close to 40% of a 200k context window spent before the agent has done anything but look.
This article uses one baseline throughout: 2,691 tokens for a 1080p frame on Claude 4.7 and later, the exact rate under that tier's patch cap. A separate baseline prices a full-resolution Retina recording at a flat 4,784-token cap per frame, because those frames exceed the cap entirely. That's a different comparison, built for a different question, and this article doesn't mix the two.
What does a screenshot cost?
One 1080p screenshot costs 2,691 tokens on Claude 4.7 and later under Anthropic's patch formula; a smaller real screenshot, 1,288×811 pixels, measured at 1,334. A realistic bug report often needs several views. Four screenshots covering one page cost 10,764 tokens with zero narration, so the agent still has to guess why each one was sent.
This is the comparison that actually matters day to day. Almost nobody sends a coding agent a video; plenty of people paste screenshots. A single one costs the same 2,691 tokens (Claude 4.7+) as a single frame of the dump above (or less, at 1,334, if it's a smaller crop rather than a full-screen capture). The cost climbs fast once a bug needs context: four screenshots to cover one page run to 10,764 tokens, and every one of them arrives with no explanation attached. For the full breakdown across common resolutions, see how many tokens a screenshot costs an AI agent.
What do words cost?
A spoken review of the same screen, transcribed, costs far less than any image. A measured 15-second test transcript ran 65 tokens; a typical spoken walkthrough runs about 150, in a 100–200 token range. That's roughly 18 times fewer tokens than one screenshot. But cheap words only help the agent if they say exactly where the problem is.
This is the gap that makes the comparison interesting rather than obvious. 150 tokens against 2,691 (Claude 4.7+) for a single screenshot, against 80,730 for the frame-dumped equivalent, is not a close race. Text is cheap because language is a compact way to encode intent; an image spends tokens encoding pixels whether or not those pixels are relevant. The catch shows up immediately in practice: "the spacing looks off" and "the spacing on the checkout card is off" cost almost the same number of tokens, and only one of them is useful. Most written feedback defaults to the first kind.
Which carries the most meaning per token?
Words win on raw efficiency, but tokens aren't the only unit that matters: meaning per token is. "The button is misaligned" costs almost nothing and says almost nothing. "The submit button on the pricing page sits right of the card edge, not centered like the others" costs barely more and pinpoints the fix. Specificity, not brevity, makes words useful.
This is the real finding here, and it cuts against the instinct to just pick the cheapest format and move on. A vague sentence and a precise one are nearly the same price in tokens; the precise one just requires the writer to do a bit more work naming what and where. Most people skip that work, which is why "words are cheapest" doesn't automatically mean "words are best" in practice. A companion article covers this in more depth: how to describe a visual bug to an AI coding agent, including concrete before-and-after phrasing.
- Words are the cheapest format by a wide margin, but only when they name an exact location.
- A vague sentence and a precise one cost almost the same number of tokens, so precision is nearly free once you're already writing.
- The format that reliably works in practice is words plus one or two selected frames, not words alone and not a full frame-dump.
What's the practical answer?
For a single, obvious bug, a specific sentence is enough on its own, and it costs next to nothing. For anything visual, spacing, alignment, colour, motion, pair that sentence with one or two frames showing exactly where. That combination measures out to about 3,500 tokens typical, roughly 23 times fewer than frame-dumping the same review.
Here's the three-way comparison laid out for one real 30-second review, all three formats scored against the same session:
| Approach | Tokens for one 30-second review | What the agent actually gets |
|---|---|---|
| Video, frame-dumped (1 fps, 1080p) | 80,730 (30 × 2,691 tokens/frame, Claude 4.7+) | Every frame, no signal for which one matters |
| Screenshots (4, covering one page) | 10,764, zero narration | A few fixed views, no "why," no order |
| Words only (typical spoken review) | ~150 | Cheap, but only useful if it names an exact location |
| Words + a couple of selected frames | ~3,500 (measured typical) | Location, intent and a visual anchor, together |
The honest answer is that no single format wins every case. A one-line, unambiguous bug ("the footer copyright year is wrong") needs nothing but a precise sentence. Anything where "where" or "what it looks like" is the whole point needs a frame attached, because prose is a poor tool for describing geometry. What consistently loses is the two extremes: a full frame-dump nobody asked for, and a vague sentence with no image to anchor it.
A short checklist for keeping feedback both cheap and specific:
- Name the exact element and where it lives on the page, not just "the layout."
- State what's wrong and what you expected, in the same sentence.
- If the bug is spatial, spacing, alignment, colour, overlap, attach one frame instead of describing geometry in prose.
- Leave everything else out. A short, precise note at the right location beats a long, rambling one every time.
- Skip the full recording. If a frame-by-frame dump feels necessary, the real problem is usually that the location wasn't named yet.
Walkie is one way to generate the words-plus-frames pairing automatically, narration transcribed alongside the specific frames it points at, but the same result is achievable by hand: record nothing, write one precise sentence, and attach a single screenshot only when the bug is genuinely visual. For a ranked list of other ways to bring the number down, see 11 ways to cut Claude Code token costs.