BroadRouter

Image token calculator

Updated

In short: Image tokens depend on each model's rule, not the file size. A 1024 × 1024 image is 1,369 tokens on Claude Sonnet 5.5 (28-pixel patches), 1,229 on GPT-5.5 (32-pixel patches × 1.2), 765 on GPT-4.1 (512-pixel tiles) and 1,120 by default on Gemini 3 and 4. Drop an image or type its size to price it on 18 models.

Free, no sign-up, runs in your browser: nothing you type is sent anywhere.

Image size and volume

Only the width and height are read, in your browser. Nothing is uploaded.

Image tokens and cost by model
ModelTokens per imagePer imagePer 1,000 imagesYour images
Measure your images with $5 free credit

Image input only: your text prompt and the model's answer are billed on top. BroadNet prices are its published input prices, synced 2026-11-14; other prices are standard list prices checked 2026-11-14. Rows marked "upstream rule" use the base model's documented rule because BroadNet does not publish its image settings; "up to" rows show a documented maximum.

Report a wrong price or number (email to info@broadnet.ai, with this page's link and your inputs filled in).

How it works

  1. Type the image width and height in pixels, or drop an image on the box (or choose a file). The page reads its dimensions in your browser; nothing is uploaded.
  2. Enter how many images you will send. Optionally set the OpenAI detail level and the Gemini media resolution.
  3. Read the table: tokens per image under each model's documented rule, the cost per image, per 1,000 images and for your volume, cheapest first.
  4. For BroadNet models, check the label: "upstream rule" means the base model's documented rule; confirm with usage.prompt_tokens. Copy the link or download CSV.

The formulas

cost per image = image tokens × input price per 1M tokens / 1,000,000

Claude          tokens = ⌈w / 28⌉ × ⌈h / 28⌉, after shrinking to fit
                2,576 px and 4,784 tokens (Claude 4.7 and later)
                or 1,568 px and 1,568 tokens (older models)
OpenAI patches  tokens = ⌈ ⌈w / 32⌉ × ⌈h / 32⌉ × multiplier ⌉, after fitting
                the detail level's size limit and patch budget
OpenAI tiles    tokens = base + tile tokens × number of 512 px tiles, after
                fitting 2048 × 2048 and scaling the short side to 768
Gemini 3 and 4  tokens = 1,120 per image (media_resolution default)
Qwen (upstream) tokens = w' × h' / (32 × 32) + 2, sides rounded to 32 px,
                at most 2,621,440 pixels
Gemma 4         tokens ≤ token budget: 70, 140, 280, 560 or 1,120
DeepSeek        tokens ≤ 1,024 (large images scaled to about 1300 × 1300)

Each rule is implemented from the provider's own documentation, and the page reproduces their worked examples: OpenAI's GPT-6 Astra high-detail examples (1024 × 1024 at 1,229 tokens, 2048 × 2048 at 3,000, 4096 × 512 at 2,458), and Anthropic's 1920 × 1080 screenshot at 1,560 tokens on the standard tier and 2,691 on the high-resolution tier. For GPT-4o and GPT-4.1, applying OpenAI's tile steps to a 1024 × 1024 image gives 768 × 768, 4 tiles and 765 tokens. See vision language model and input and output tokens in the glossary.

Which rule does each model use?

ModelInput, per 1M tokensImage token rule1024 × 1024 imageSource
DeepSeek V4.1 Flash (BroadNet)$0.175DeepSeek's documented cap: up to 1,024 tokens per image. Upstream rule, not published by BroadNet. Upstream rule from DeepSeek's own API docs for deepseek-flash (DeepSeek V4.1 Flash): large images are scaled to about 1300×1300 pixels' worth, small ones scaled up. No exact formula is published, so this is an upper bound.up to 1,024 tokensDocs
Qwen3.8 27B (BroadNet)$0.17532×32 px per token + 2, at most 2,621,440 pixels (2,560 tokens). Upstream rule, not published by BroadNet. Upstream rule from Alibaba Cloud Model Studio for Qwen3.8 and Qwen3-VL (default max_pixels). BroadNet's open-weight build may use other limits.1,026 tokensDocs
BroadNet Gemma 4 VL 26B (BroadNet)$0.10Gemma 4 token budget: 70, 140, 280, 560 or 1,120 per image. Upstream rule, not published by BroadNet. Upstream rule from Google's Gemma 4 docs: the budget is a hard limit per image, whatever its size. BroadNet does not state which budget its build uses, so the range is shown.70 to 1,120 tokensDocs
GPT-6 Astra (OpenAI)USD 10.0032 px patches × 1.2; high: 2,500-patch budget; auto: original size. Price for prompts up to 272K tokens.1,229 tokensDocs, checked 2026-11-14
GPT-5.6 Terra (OpenAI)USD 2.0032 px patches × 1.2; high: 2048 px and 2,500 patches; auto: original size. Price for prompts up to 272K tokens.1,229 tokensDocs, checked 2026-11-14
GPT-5.6 Luna (OpenAI)USD 0.2032 px patches × 1.2; high: 2048 px and 2,500 patches; auto: original size. Price for prompts up to 272K tokens.1,229 tokensDocs, checked 2026-11-14
GPT-5.5 (OpenAI)USD 5.0032 px patches × 1.2; high: 2048 px and 2,500 patches; auto: 6000 px and 10,000 patches. Price for prompts under 272K tokens.1,229 tokensDocs, checked 2026-11-14
GPT-5.4 mini (OpenAI)USD 0.7532 px patches × 1.2; auto and high: 2048 px and 2,500 patches; low: 6,144 patches. 1,229 tokensDocs, checked 2026-11-14
GPT-4.1 (OpenAI)USD 2.0085 + 170 per 512 px tile, after fitting 2048 px and a 768 px short side. 765 tokensDocs, checked 2026-11-14
GPT-4o (OpenAI)USD 2.5085 + 170 per 512 px tile, after fitting 2048 px and a 768 px short side. 765 tokensDocs, checked 2026-11-14
Claude Opus 5.5 (Anthropic)USD 4.00⌈w/28⌉ × ⌈h/28⌉; high-resolution tier: up to 2,576 px and 4,784 tokens. 1,369 tokensDocs, checked 2026-11-14
Claude Sonnet 5.5 (Anthropic)USD 2.00⌈w/28⌉ × ⌈h/28⌉; high-resolution tier: up to 2,576 px and 4,784 tokens. 1,369 tokensDocs, checked 2026-11-14
Claude Haiku 5.5 (Anthropic)USD 0.10⌈w/28⌉ × ⌈h/28⌉; high-resolution tier: up to 2,576 px and 4,784 tokens. Price for prompts up to 100,000 tokens (USD 0.50 above).1,369 tokensDocs, checked 2026-11-14
Claude Haiku 4.5 (Anthropic)USD 1.00⌈w/28⌉ × ⌈h/28⌉; standard tier: up to 1,568 px and 1,568 tokens. 1,369 tokensDocs, checked 2026-11-14
Gemini 4 Hydrogen (Google)USD 1.50Fixed per image by media_resolution (default 1,120). 1,120 tokensDocs, checked 2026-11-14
Gemini 3.1 Pro Preview (Google)USD 2.00Fixed per image by media_resolution (default 1,120). Price for prompts up to 200K tokens.1,120 tokensDocs, checked 2026-11-14
Gemini 3.8 Flash (Google)USD 0.75Fixed per image by media_resolution (default 1,120). Price through 2026-12-31; Google lists USD 1.50 from 2027-01-01.1,120 tokensDocs, checked 2026-11-14
Gemini 3.5 Flash-Lite (Google)USD 0.30Fixed per image by media_resolution (default 1,120). 1,120 tokensDocs, checked 2026-11-14

Prices are standard per-token input prices; batch and cached prices are lower. On BroadNet Inference there is no separate image fee: image tokens bill at the input price shown on the pricing page.

Worked example: 10,000 scanned invoices

A phone photo of an invoice at 2000 × 1500 pixels is 3,888 tokens on Claude Sonnet 5.5 and 3,554 on GPT-5.5 at auto detail. On Qwen3.8 27B, the upstream Qwen rule gives 2,496 tokens: USD 0.44 per 1,000 images at $0.175 per 1M input tokens. On BroadNet Gemma 4 VL 26B, even the largest Gemma 4 budget of 1,120 tokens costs USD 0.11 per 1,000 images. DeepSeek V4.1 Flash is capped at 1,024 tokens per image by DeepSeek's rule, at most USD 0.18 per 1,000 images. Add your prompt and the extracted JSON, which often cost more than the image. The LLM API cost calculator prices the full request.

How to cut image token costs

  • Resize before you send. Above each model's cap you pay for pixels the model throws away anyway.
  • Crop to the region you need: a receipt total, a table, one form field.
  • Use a low detail level or low media resolution for classification and captions; keep high detail for small text.
  • Measure once: compare usage.prompt_tokens with and without the image on 20 real samples, and use the token counter for the text part.

Questions

How many tokens is an image?

It depends on the model's rule and the image size, not the file size. A 1024 × 1024 image is 1,369 tokens on Claude Sonnet 5.5, 1,229 on GPT-5.5, 765 on GPT-4.1 and 1,120 by default on Gemini 3 and 4 models, using each provider's documented rule (checked 2026-11-14). Image tokens are billed at the model's normal input price.

How many tokens is a 1024 × 1024 image in GPT?

On GPT-5.5 and the GPT-5.6 models, OpenAI counts 32 × 32-pixel patches and multiplies by 1.2: 1,024 patches, so 1,229 tokens. On GPT-4o and GPT-4.1 it uses 512-pixel tiles: the image is scaled to 768 × 768, which is 4 tiles, so 85 + 4 × 170 = 765 tokens. With detail set to low, GPT-4.1 charges only the 85 base tokens.

How does Claude count image tokens?

Anthropic's vision docs (checked 2026-11-14) count 28 × 28-pixel patches: tokens = ⌈width ÷ 28⌉ × ⌈height ÷ 28⌉. Larger images are first shrunk to fit the model's limits: 2,576 pixels and 4,784 tokens on Claude 4.7 and later, 1,568 pixels and 1,568 tokens on older models. The older width × height ÷ 750 rule of thumb is only an approximation: it overestimates large images and underestimates small ones (a 200 × 200 image is 64 tokens, not 53).

How many tokens does Gemini use per image?

On Gemini 3 and Gemini 4 models the count is set by the media_resolution setting, whatever the image size. Google's table lists approximately 1,120 tokens by default and at high, 560 at medium, 280 at low and 2,240 at ultra_high (per content item only). Google's token-counting page also gives a general rule: 258 tokens for images up to 384 pixels on both sides, and 258 per 768 × 768 tile above that.

Does resizing an image reduce its token cost?

For patch-based models (Claude, GPT-5.x, Qwen), yes: fewer pixels means fewer patches, up to each model's cap. For fixed-budget models it does not: Gemini 3 and 4 use the media_resolution count, Gemma 4 uses its token budget, and DeepSeek scales every large image to about the same size. Resize to the smallest size that still reads your text; very small text needs more pixels.

How do I measure image tokens on BroadNet Inference?

BroadNet does not publish image-token settings for its builds, so this tool shows the upstream model's documented rule. To measure, send the same prompt once with and once without the image and subtract the two usage.prompt_tokens values. On BroadNet, image tokens are billed as input tokens at the model's input price. How to count tokens before sending.

Go deeper

Read as MarkdownAsk ChatGPTAsk ClaudeAsk Perplexity