Chandra 2.6 is live in the Datalab API, and every mode (fast, balanced and accurate) runs on it. You don’t need to change anything to use it, and the price is the same.
What’s new:
- More accurate than frontier models from Google, Anthropic and OpenAI. On OmniParseBench, Datalab accurate scores 94.1, against 92.7 for Gemini 3.8 Flash, 91.4 for GPT-6.1 Sol and 90.7 for Claude Opus 5.5.
- 2.5x faster than Chandra OCR 2.4. On a fixed set of 1,187 OmniParseBench pages, the median time per page went from 17.0 s to 6.7 s.
- More reliable. In our logs, failed requests went from 1 in 6,400 with Chandra 2.5 to 1 in 112,000, and slow pages are much rarer than with Chandra OCR 2.4.
Faster
To compare releases fairly, we ran each one on the same 1,187 OmniParseBench pages, on the same GPU:
Chandra 2.6 is 2.5x faster than Chandra OCR 2.4 at the median, and 37% faster than Chandra 2.5. These times are for the model on our GPUs, without upload and queueing, and include today’s serving update.
API users see the change too. On the 200 OmniParseBench pages we use for the benchmark’s latency test, sent through the API 8 at a time, the median time per page went from 9.6 s with Chandra 2.5 to 6.0 s in balanced mode, and from 14.8 s to 11.4 s in accurate mode.
In production, for the same customers in every period, here is the median time they waited for a single page, end to end:
API latency test: the same 200 pages and settings on October 9 (Chandra 2.5) and today, after the serving update (Chandra 2.6). Fixed pages: the same 1,187 pages, the mean of several runs per release. Production: single-page API requests, end to end, for the same customers in every period. Chandra OCR 2.4: September 30 to October 6, 2026. Chandra 2.5: October 15 to 23. Chandra 2.6: October 23 to 27.
More reliable
Fewer requests fail, and fewer pages are slow.
In our logs, 1 in 6,400 single-page API requests failed in the last days of Chandra 2.5. Since the Chandra 2.6 release, that’s down to 1 in 112,000, and every failure was an input image that couldn’t be opened.
Slow pages are rarer. On the fixed pages, the p90 went from 33.3 s with Chandra OCR 2.4 to 14.1 s. In production, the p90 is 68% lower in balanced mode and 77% lower in accurate mode.
Single-page /convert and /marker requests. Chandra 2.5: October 21 to 23, 2026. Chandra 2.6: October 23 to 27.
More accurate
Chandra 2.6 tops the published OmniParseBench results, ahead of frontier models from Google, Anthropic and OpenAI. OmniParseBench is our open parsing benchmark with 16,288 pass/fail tests on 2,937 pages in 92 languages. Datalab accurate scores 94.1, up from 92.7 with Chandra 2.5, and ahead of Gemini 3.8 Flash at 92.7. Datalab balanced scores 92.6, up from 91.6, which is also higher than GPT-6.1 Sol (91.4) and Claude Opus 5.5 (90.7).
On hard pages, accurate mode scores higher than Gemini 3.8 Flash on tables (95.6 vs 93.8), multi-column pages (94.9 vs 92.3), rotated pages (95.1 vs 92.5) and degraded scans (89.7 vs 85.4).
* Used in creating and verifying at least some tests. Datalab scores are a rerun of the production API on October 23, 2026, on the same 16,288 tests. † Our runs on October 27, 2026, with the benchmark’s standard model adapter. Pages a model refused to answer count as failures in its overall score. Its family scores cover the pages it answered. Full results, including every hard slice, are in the OmniParseBench results.
Same price
The price stays the same: fast and balanced are $4 per 1,000 pages, and accurate is $10. See pricing.
Try it
Chandra 2.6 is live in the API and the playground. For on-premises deployments, contact us at [email protected].
If you have hard documents you want us to try, send them to [email protected], or find me on Twitter.
