Best AI Reranker Models for RAG, Ranked by Retrieval Quality and Cost

Voyage rerank-2.5 is the quality-optimized tier of Voyage AI's 2.5 reranker series, released in August 2025 and distributed through Voyage's API, AWS Marketplace, and MongoDB Atlas. On a standard suite of 93 retrieval datasets spanning multiple domains, rerank-2.5 and rerank-2.5-lite improve retrieval accuracy by 7.94% and 7.16% over Cohere Rerank v3.5. The model was the first in the rerank family to ship instruction-following, which lets users steer relevance scores with natural language, and on the Massive Instructed Retrieval Benchmark (MAIR), rerank-2.5 and rerank-2.5-lite outscore Cohere Rerank v3.5 by 12.70% and 10.36%, respectively. The trade-offs are throughput and ecosystem reach: latency is 1.5s for 25K tokens, throughput is 60M tokens per hour at $0.05 per 1M tokens on an ml.g6.xlarge, and per-call latency on short candidates trails Cohere Rerank 3.5.

Cohere Rerank 4 Pro is Cohere's current state-of-the-art quality tier and the broadest production deployment in this list, with Bedrock, Azure Marketplace, and OCI distribution behind it. On the 15 February 2026 Agentset leaderboard snapshot, it sits second on head-to-head ELO at 1629, behind Zerank-2 at 1638 and ahead of Voyage Rerank-2.5 at 1544, though Agentset notes leaderboards in this cluster shift with each release. Pricing is per-search rather than per-token: Rerank 4 Pro at $0.0025 each, Rerank 4 Fast at $0.002, and Rerank v3.5 at $0.001 (or $2.00 per 1,000 queries on Bedrock), with one 'search' defined as a query plus up to 100 documents. Dedicated capacity runs through Model Vault, from $4.00/hour ($2,500/month) for an Embed 4 Small instance up to $10.00/hour ($6,500/month) for a Rerank 4 Pro Large instance. It's the right pick when multi-cloud procurement and a predictable per-search line item outweigh the last point of nDCG.

Jina Reranker v3 is the open-weight quality leader in this comparison. It's a 0.6B-parameter multilingual document reranker built on a 'last but not late' interaction architecture: unlike ColBERT's separate encoding with multi-vector matching, the model runs causal self-attention between query and documents inside the same context window, then extracts contextual embeddings from the last token of each document. Built on Qwen3-0.6B with 28 transformer layers and a lightweight MLP projector, it processes up to 64 documents simultaneously within a 131K-token context and posts state-of-the-art BEIR performance at 61.94 nDCG@10, 10× smaller than generative listwise rerankers. It outperforms Qwen3-Reranker-4B at 6× smaller size, and on multilingual retrieval reaches 66.50 on MIRACL across 18 languages, with Arabic at 78.69 and Thai at 81.06. The trade-off is operational: it's self-hosted, so the cost line shifts from per-call API spend to amortized GPU infrastructure, and the weights ship under CC BY-NC 4.0, which restricts direct commercial use.

Cohere Rerank 3.5 is the previous-generation Cohere tier still in wide production deployment in 2026, largely on the strength of its price point and language coverage. Rerank v3.5 handles English and multilingual documents as well as semi-structured data like JSON, with a 4,096-token context. Long documents are automatically chunked and the highest relevance score among chunks is used for ranking. Cohere reports leading performance across over 100 languages, and on AWS Bedrock, Rerank v3.5 is priced at $0.001 per search ($2.00 per 1,000 queries on Bedrock). The trade-offs are quality and context: it trails Voyage rerank-2.5 by 7.94% on Voyage's 93-dataset benchmark and trails Rerank 4 Pro on Cohere's own quality positioning, and the 4,096-token chunk limit forces truncation on long documents. On Agentset's tests, Cohere v3.5 was the fastest in the field but less favored by the LLM judge, while Voyage 2.5 hit the middle ground.

BGE Reranker v2-m3 is the long-standing open-source baseline that newer rerankers get measured against in 2026. It's lightweight, multilingual, easy to deploy, and fast at inference. The practical test is whether a newer model outperforms it by enough margin to justify the added cost or latency. Jina's published comparison puts jina-reranker-v3 5.43% above same-scale bge-reranker-v2-m3 on BEIR, and on Agentset's leaderboard BGE v2-M3 shows sharp spikes, performing well only in select cases. It's the right pick when Apache-2.0 weights and full deployment control are the binding constraints; for hosted workloads, the 2025-2026 models post measurably better numbers.