Your JPEG Quality Knob Saves 0%
Cody Lee Walker · codylee.ca · 2026-07
Here is a measurement that surprises almost everyone who ships images to a vision-LLM: we took real screenshots, requantized them to JPEG quality 25 — visibly degraded, files a fraction of the size — and measured the token bill. It moved by exactly 0.0%.
Not roughly zero. Exactly zero, because vision-LLM billing never looks at your bytes. Each provider converts pixel dimensions to tokens with its own arithmetic:
| provider | what you pay for |
|---|---|
| Anthropic | ~(w·h)/750 after a 1568px long-edge fit and a ~1.15-megapixel cap |
| OpenAI (high detail) | 85 base + 170 per 512px tile, after a 2048px square fit and 768px short-edge fit |
| Gemini | 258 per 768px tile (flat 258 when both sides ≤ 384px) |
A 2560×1440 screenshot bills ~1,533 Anthropic tokens whether it's a 4MB PNG or a 90KB mud-quality JPEG. The quality slider — the knob every image CDN, every "optimize for web" preset, every compression SaaS puts in front of you — optimizes a number the model provider doesn't bill on.
Two practical consequences:
1. The only lever is downscaling. Fewer pixels, fewer tokens — same 2560×1440 screenshot at 827×465 bills ~513 tokens, a 67% cut. Everything else is placebo. (Format still matters for correctness — some providers re-encode PNGs, and a deterministic re-encode keeps prompt caches warm — but not for the bill.)
2. Downscaling has a cliff, and the cliff moves. Shrink a chart until the axis labels blur and the model doesn't degrade gracefully — it reads confidently wrong. Where that cliff sits depends on the content and the model reading it, which is why "resize everything to 1024" is a policy that quietly fails on your densest, most valuable images first.
So the real question was never "what quality setting" — it's "how far can this image shrink before this model misreads it?" That's a per-image, per-model measurement. It's exactly the measurement a starlens certificate performs, seals, and prices: you pay a share of the tokens it proves it saved, and when no safe compression exists you pay nothing.
Free first image at api.s2ar.dev/app —
or estimate your current bill locally, no key needed, with
starlens estimate screenshot.png.
Compress your first image free
starlens shrinks what you send a vision-LLM and hands back a sealed, replayable certificate that the model still reads it the same — or a certified refusal, which costs nothing. Pricing is a share of the savings the certificate proves; you keep the rest.