Fast mode now runs Surya 2.6, our new OCR model. For every fast-mode request:
- It costs $1 per 1,000 pages when you share data for training ($1.33 without), down from $3 ($4 without).
- It scores 91.6 on OmniParseBench, up from 90.3. Tables gain the most, from 88.0 to 91.2.
- It’s faster. The model returns a page in 2.6 seconds at the median, down from 3.9, and one GPU reads 1.7 times as many pages.
Nothing changes in your code. Every request with mode="fast" uses Surya 2.6 from today.
Where fast mode got better
| OmniParseBench | Fast mode before | Fast mode now |
|---|---|---|
| Overall | 90.3 | 91.6 |
| Text | 87.0 | 87.8 |
| Tables | 88.0 | 91.2 |
| Layout | 96.0 | 95.9 |
| Old scans | 79.1 | 75.2 |
Tables gained 3.2 points. The largest gains are on pages that used to trip fast mode up:
- Long pages of tiny text went from 84.8 to 95.8.
- Rotated and skewed pages went from 85.4 to 92.0.
- arXiv pages with math went from 92.5 to 95.0.
Fast mode also gets stuck far less often. Before, 35 of the 2,937 pages ran until they hit the output length limit. Now 2 do.
Where to use balanced or accurate instead
Fast mode is weaker on some documents than before:
- Handwriting fell on two test sets: from 69.1 to 54.3, and on French handwriting from 46.4 to 45.2.
- Old scanned text fell from 72.2 to 63.0, and old scans overall fell 3.9 points.
- Very large pages fell from 86.0 to 80.7.
- Filled-in forms fell from 88.8 to 86.0, and a set of real-world PDFs from 95.3 to 93.4.
If your documents are handwritten, old scans or large-format pages, use balanced or accurate.
Not sure which mode fits your documents? Run fast, balanced and accurate side by side on your own pages with Evals. It shows every place the modes read the page differently.
Faster
On the same GPU, with 16 requests in flight on 1,187 pages, the model’s median time per page fell from 3.87 seconds to 2.58, 33% less. The slow pages got faster too: the 90th percentile fell from 9.15 seconds to 5.20. One GPU now reads 1.66 times as many pages per second.
These are model times. The time through the API also includes the upload, rendering and the rest of the pipeline.
$1 per 1,000 pages
| Mode | With data sharing | Without |
|---|---|---|
| Lite | $0.38 | $0.50 |
| Fast | $1 (was $3) | $1.33 (was $4) |
| Balanced | $3 | $4 |
| Accurate | $7.50 | $10 |
Sharing your documents for model training takes 25% off every mode. Fast mode now gives you full layout, tables, math and reading order for a third of its old price.
Try fast mode in the playground.
How we measured
Both runs used OmniParseBench, all 2,937 pages, with the production prompt. “Before” is fast mode as it ran until today, at the same page resolution it used in production. Speed: one GPU, 1,187 pages, 16 requests in flight, each model with its production settings. Throughput comes from one run per model, so read it as plus or minus 5%.