Open OCR costs fell 5x in 18 months: $0.51 to $0.095 per 1,000 pages, benchmarked on 850 pages
vanstriendaniel · x · 2026-11-13
The author benchmarked 8 open OCR models from the last 18 months on the same 850 pages on one Hugging Face Jobs GPU.
Key findings
- Cost per 1,000 pages dropped from $0.51 to $0.095, roughly halving every 10 months—but most of the drop comes from one model: LightOnOCR-3-0.8B, about 5x cheaper.
- How you run a model matters too: GLM-OCR on 250 dense pages went from $1.57/1k pages (February setup) to $0.40 today (many pages in flight, pages fitted to 2,048 px, newer vLLM, bigger GPU).
- Full-size scans didn't help: 1.3–4.2x more expensive than 2,048 px with no accuracy gain on book pages.
- On 282 checked book pages, every model gets its median page over 99% right; two cheaper models make 1.5x the errors of the oldest.
- On dense two-column 1771 print, the oldest model beat most newer ones; the newest only ties it.
Testing all eight on 850 pages cost about $6 of GPU time—still worth testing on your own pages. A full post is coming.
Related event: Open-Source OCR Cost Drops 5x in 18 Months, Benchmark Shows(2 posts)→
More from Infra
- llama.cpp ships v1.0.0 after 3.5 years, committing to semver API/ABI stability for local AI — ggerganov · 2026-11-13
- RTX 5070 Ti vs RTX 3090 in ComfyUI: 5070 Ti wins warm batches, 3090 wins cold loads — Numerous_Access_8008 · 2026-11-13
- Futurum: NVIDIA's AI factory blueprint sets the standard, agent software at ~$2,000 per GPU per year — lifebypixels · 2026-11-13
- llama.cpp hits v1.0 as Hugging Face ships one-click Llama desktop apps for local AI — unofficialmerve · 2026-11-13
- llama.cpp ships v1.0.0, a milestone foundation for the local AI ecosystem — Hugging Face Blog · 2026-11-13
- Microsoft outage sparks warning: a million AI agents can share one point of failure — tallmetommy · 2026-11-13