Claude image token calculator
Estimate how many input tokens an image costs on Claude (Sonnet / Opus / Fable), and whether packaging your prompt as an image beats sending it as text.
| Model family | Image tokens (est.) | vs 2,000 text tokens |
|---|---|---|
| Claude | ~324 | 84% fewer tokens |
Formulas: Claude — ≈ one token per 28×28-pixel visual patch (rounded up per side). Estimates only, before any wrapper instructions or caching — verify with the provider's token-counting API. See our methodology.
Common sizes on Claude
| Square image | Estimated tokens | Beats text longer than |
|---|---|---|
| 384×384 | ~196 | ~196 text tokens |
| 500×500 | ~324 | ~324 text tokens |
| 512×512 | ~361 | ~361 text tokens |
| 768×768 | ~784 | ~784 text tokens |
| 1024×1024 | ~1,369 | ~1,369 text tokens |
Notes and caveats
Anthropic downscales very large images before processing, so pushing the canvas size up does not increase capacity indefinitely. Multiple smaller pages at a readable font usually beat one dense oversized page.
If you repeatedly send the same context to Claude, compare image packaging against prompt caching first: caching is lossless and discounts repeated prefixes, while image compression is lossy. For Claude Code and agent workflows, compress the static background material and keep live state as text.
These are input-side estimates before wrapper instructions, caching, latency, retries and accuracy effects. Token rules are model-version-specific and change — verify with the provider's token-counting endpoint, and read when not to convert prompts to images before shipping this in production.
Try it on a real prompt
The visual prompt compressor gives you a live Claude-specific estimate, flags exact identifiers that shouldn't go into a lossy image, and exports the PNG plus the original .txt — all locally in your browser.
Last updated July 15, 2026 · Estimates are derived from provider documentation and can change — see methodology.