Anthropic Claude

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.

Claude models process images as visual patches of roughly 28×28 pixels, so a 500×500 image works out to about 18 × 18 = 324 tokens. A prompt image starts saving tokens on Claude once the text it replaces exceeds roughly 324 tokens (~1,300 characters) — and only if the rendered text stays legible at that density.
Model familyImage tokens (est.)vs 2,000 text tokens
Claude~32484% 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 imageEstimated tokensBeats 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.