Interactive · Week 10 · Section 2

One image is roughly 1,000 tokens.

Set the multimodal workload — model, image count, resolution, traffic. The calculator shows how image tokens dominate the cost equation, and how the same workload's bill can vary 50× across vision-capable models. The visceral lesson: adding 5 high-res images to each request typically multiplies your bill by 5-10×. Most teams don't do this math before shipping vision features and discover the cost only when the bill arrives.

Workload

Vision-capable model —

Requests per second 5.0

—

Resolution tier scales token cost per image roughly 18× from low to high

Per-request shape

Images per request 3

3 images × 768 tokens = 2,304 image tokens

Text input tokens 500

prompt + system prompt + retrieved text

Output tokens 300

model's text response

Per request

—

Per year

—

at current RPS

Text-only equivalent

—

/year · 0 images

Image bill multiplier

—

vs text-only baseline

Annual cost · text-only vs multimodal

—

Text only — / year

With images — / year

Cost breakdown · per request

—

image tokens

text input

output

Same workload, all vision-capable models

cheapest highlighted · current model has accent border

What's happening here

Image token counts are normalized across models for comparison: low resolution ≈ 85 tokens, medium ≈ 768 tokens, high ≈ 1568 tokens per image. Real provider tokenization differs slightly (GPT uses 765 + 170 per 512px tile; Gemini uses 258 base; Claude uses 85/768/1568 tiers) but the order-of-magnitude story is the same. Pricing reflects 2026 hosted-API rates. Source-of-truth lives as <MultimodalCostExplorer/> in the curriculum codebase.