Skip to content

Cost & timing

Inspect the cost and processing time of scientific knowledge extraction across the 74-page Distils showcase. These measurements cover PDF processing, fact extraction, and entity selection; they exclude downstream RAG indexing or GraphRAG construction. They are archived observations, not a current price quote or a general model benchmark.

Measurement conditions

Both runs used the OrcaRouter API, one worker, and DeepSeek in non-thinking mode. No sampling parameter other than temperature was set.

On September 13, 2026, OrcaRouter listed the following DeepSeek V4.1 Flash prices:

UsageOff-peakWeekday peak
Uncached input / 1M tokens$0.15$0.30
Cached input / 1M tokens$0.003$0.006
Output / 1M tokens$0.60$1.20

Weekday peak windows were 01:00-04:00 and 06:00-10:00 UTC, with every rate doubled. Both archived runs were off-peak: they ran on Sunday from 10:50:16 to 10:59:30 UTC and from 10:59:41 to 11:07:58 UTC.

API cost

ModelTotal API costMean / pageMedian (P50) / pageP95 / page
DeepSeek V4.1 Flash, T=0$0.116260$0.001571$0.001617$0.003033
DeepSeek V4.1 Flash, T=1$0.099126$0.001340$0.001429$0.002103

Processing time

ModelTotal pipelineMean API time / pageMedian (P50) / pageP95 / page
DeepSeek V4.1 Flash, T=09.8 min7.33 s7.32 s12.50 s
DeepSeek V4.1 Flash, T=18.4 min6.55 s6.86 s10.52 s

Per-page API values combine extraction and entity calls, with document selection distributed evenly across the paper's pages. Pipeline time includes fresh GROBID processing and orchestration; PDF rendering is recorded separately. Cost is reported by OrcaRouter.

Request-level statistics

API cost

ModelTotal API costMean / requestMedian (P50) / requestP95 / request
DeepSeek V4.1 Flash, T=0$0.116260$0.000807$0.000659$0.001838
DeepSeek V4.1 Flash, T=1$0.099126$0.000684$0.000556$0.001492

Processing time

ModelTotal pipelineMean API time / requestMedian (P50) / requestP95 / request
DeepSeek V4.1 Flash, T=09.8 min3.77 s3.34 s6.70 s
DeepSeek V4.1 Flash, T=18.4 min3.34 s3.08 s6.44 s

P50 is the median; P95 uses the nearest rank. All received responses, including retries, contribute to the measurements. A single run per configuration does not establish a general speed or quality ranking.

Released under the Apache License 2.0.