Voyage rerank-3
Voyage AI’s hosted reranker family, launched 2026-09-30 in two sizes, rerank-3 and the latency-optimized rerank-3-lite (launch post). It succeeds rerank-2.5 and rerank-2.5-lite (August 2025) as a drop-in upgrade with the same API and price.
Properties
| rerank-3 | rerank-3-lite | |
|---|---|---|
| Price | $0.05 / M tokens | $0.02 / M tokens |
| Context | 32K tokens per document | 32K tokens per document |
| Instruction-following | yes | yes |
- Score calibration is preserved across versions. Scores match the distributions of the corresponding 2.5 models, so existing thresholds keep working (Calibrated Relevance Probability).
- Positioning: largest gains on long documents and code, attributed to coding agents issuing more queries and to longer candidates.
- Availability: Voyage API (first 200M tokens free), and natively in MongoDB Atlas Vector Search.
Evidence
Vendor-reported, averaged across four first-stage retrievers on 95 datasets: +2.72% NDCG@10 over Cohere Rerank v4.0 Pro and +3.02% over Qwen3-Reranker-8B. On LongEmbed the lead over Cohere Pro is +13.84%.
Independent: Hev meets Jev measured “Voyage rerank-3” at 0.504 mean nDCG@10 on three BEIR subsets over a BM25 top-30. That is second only to Mixedbread’s mxbai-rerank-v3.1-listwise (0.511), and the cheapest of the hosted rerankers at $0.50 per 1,000 queries, with p50/p95 latency of 185/287 ms. That post predates the official launch by two weeks.
Versus Jev
Two published comparisons point in the same direction by different margins. In Hev’s BEIR run, untuned Jev trailed rerank-3 by 0.003 nDCG@10, with a much wider latency tail. Voyage’s own run on 26 expertise-heavy datasets had rerank-3 ahead by 1.77 (code), 5.69 (law) and 5.87 (finance) NDCG@10 points. That run used untuned Jev and was cut short by rate limits.