Voyage rerank-3 / 3-lite + Atlas Native `$rerank` (Preview)

Same-day dual primary (2026-09-30): Voyage ships rerank-3 and rerank-3-lite—available today per the Voyage blog (do not invent GA)—and MongoDB Atlas adds Native Reranking via aggregation-stage $rerank, documented as a Preview feature so agents can rescore candidates in-pipeline without a separate vendor hop (Voyage, MongoDB blog, docs).

This is a Desk Bot dev/rag briefing. Fence from D7 Atlas Agent Engine (runtime/memory/governance)—this is the retrieval-quality / pipeline rerank layer. Do not frame either way as a substitute for Atlas Vector Search; $rerank sits after vector/search/fusion stages.

HARD status split

Surface	Status language
Voyage rerank-3 / rerank-3-lite	Blog: available today — paraphrase that; never invent “GA”
Atlas Native $rerank / Native Reranking	Docs: Preview throughout (banner + model note); not production-recommended per docs
Mongo product blog “$rerank, available today”	Still carry the Preview label from docs

Voyage models

Model IDs: rerank-3 · rerank-3-lite. Same API / 32K context / instruction-following / price as the Rerank 2.5 series; score distributions calibrated so 2.5 thresholds still work—drop-in model-name upgrade (Voyage).

Pricing (Voyage primary): rerank-3 $0.05 / rerank-3-lite $0.02 per 1M tokens (same as 2.5 / 2.5-lite). Lite claimed ~on-par with rerank-2.5 quality at 40% of that price—attribute Voyage; not desk-verified.

Hook domains (vendor): largest claimed gains on long documents and code; 32K token context retained.

Atlas Native $rerank (Preview)

Index-less aggregation stage after $vectorSearch / $search / $rankFusion / $scoreFusion; numDocsToRerank ≤ 1000; enable in Project Settings (docs).

Docs: rerank-3 / rerank-3-lite on Atlas are Preview and need a dedicated cluster (M10+) on MongoDB 9.0+. Also available via Atlas Embedding and Reranking API / Voyage API (docs, MongoDB blog).

Soft: vendor lifts (attribute)

Any lifts vs Cohere Rerank v4.0 Pro / Qwen3-Reranker-8B (and LongEmbed / code / MAIR % figures) are Voyage/MongoDB claims—not desk-verified. Soft-attribute only (Voyage, MongoDB blog).

Who should care

RAG teams that want second-stage rerank atop embeddings/lexical/hybrid—no re-index—should start at the Voyage rerank-3 post and Atlas $rerank docs—keep Voyage “available today” ≠ GA, keep Atlas Preview + M10+/9.0+ gates, soft-attribute every bench, and leave Agent Engine to D7.