Rerank 5 Fast

Changes from previous version

Same-day latency tier under Rerank 5 Pro. Token price is $0.02 per million instead of $0.05. RCP-nDCG@10 moves from 92.2 to 90.6 on ViDoRe V3, 90.4 to 88.1 on BEIR, 82.2 to 77.5 on BRIGHT, and 87.3 to 85.1 on nanoMTEB. Throughput on one A100 is 187 documents per second against 65 for Pro.

Release Summary

Latency tier of Rerank 5, API id rerank-v5.0-fast, shipped the same day as Rerank 5 Pro. List price is $0.02 per million tokens, with the same 32,768-token context. RCP-nDCG@10 is 90.6 on ViDoRe V3, 88.1 on BEIR, 77.5 on BRIGHT, and 85.1 on nanoMTEB, above Rerank 4 Pro on each. Parsed-PDF retrieval averages 88.5. On one A100, with 96 documents per request, Fast reranks 187 standard 1,000-token documents per second and Pro reranks 65. Median latency is 4.1 seconds for Fast and 11.7 seconds for Pro. On the six-language suite, Fast scores 88.6 Spanish, 88.6 French, 88.6 German, 86.5 Arabic, 86.8 Japanese, and 89.9 Korean.

Timeline

Other Rerank versions