September 2026 Summaries

Voyage has introduced rerank-3 and rerank-3-lite, updated reranking models designed to reorder search candidates more accurately in retrieval pipelines, particularly for long documents and code-related queries. The models retain the prior Rerank 2.5 series’ API compatibility, 32K-token context window, instruction-following support, calibrated scoring behavior, and pricing, enabling existing users to upgrade without code changes. Across 95 datasets spanning nine domains and four first-stage retrieval methods, Voyage reports that rerank-3 and rerank-3-lite outperform their predecessors and several competing rerankers, including Cohere Rerank v4.0 Pro and Qwen3-Reranker-8B, with particularly large reported gains on long-document retrieval. Rerank-3-lite is positioned as a lower-latency option that approximately matches rerank-2.5’s quality at 40% of its price, while both new models cost the same as their corresponding 2.5 versions, at $0.05 and $0.02 per million tokens respectively. The models are available through Voyage’s API and MongoDB Atlas integrations, while the older Rerank 2.5 series remains available for current users.