Cohere Rerank 3.5 vs Cohere Rerank 4 Pro

Detailed comparison between Cohere Rerank 3.5 and Cohere Rerank 4 Pro. See which reranker best meets your accuracy and performance needs.

Model Comparison

Cohere Rerank 4 Pro takes the lead.

Both Cohere Rerank 3.5 and Cohere Rerank 4 Pro are powerful reranking models designed to improve retrieval quality in RAG applications. However, their performance characteristics differ in important ways.

Why Cohere Rerank 4 Pro:

  • Cohere Rerank 4 Pro has 175 higher ELO rating
  • Cohere Rerank 4 Pro delivers better accuracy (nDCG@10: 0.219 vs 0.200)
  • Cohere Rerank 4 Pro has a 17.2% higher win rate

Overview

Key metrics

ELO Rating

Overall ranking quality

Cohere Rerank 3.5

1452

Cohere Rerank 4 Pro

1627

Win Rate

Head-to-head performance

Cohere Rerank 3.5

41.0%

Cohere Rerank 4 Pro

58.2%

Accuracy (nDCG@10)

Ranking quality metric

Cohere Rerank 3.5

0.200

Cohere Rerank 4 Pro

0.219

Average Latency

Response time

Cohere Rerank 3.5

392ms

Cohere Rerank 4 Pro

614ms

Visual Performance Analysis

Performance

ELO Rating Comparison

Win/Loss/Tie Breakdown

Accuracy Across Datasets (nDCG@10)

Latency Distribution (ms)

Breakdown

How the models stack up

MetricCohere Rerank 3.5Cohere Rerank 4 ProDescription
Overall Performance
ELO Rating
1452
1627
Overall ranking quality based on pairwise comparisons
Win Rate
41.0%
58.2%
Percentage of comparisons won against other models
Pricing & Availability
Price per 1M tokens
$0.050
$0.050
Cost per million tokens processed
Release Date
2024-12-02
2025-12-11
Model release date
Accuracy Metrics
Avg nDCG@10
0.200
0.219
Normalized discounted cumulative gain at position 10
Performance Metrics
Avg Latency
392ms
614ms
Average response time across all datasets

Dataset Performance

By field

Comprehensive comparison of accuracy metrics (nDCG, Recall) and latency percentiles for each benchmark dataset.

MSMARCO

MetricCohere Rerank 3.5Cohere Rerank 4 ProDescription
Accuracy Metrics
nDCG@5
0.522
0.562
Ranking quality at top 5 results
nDCG@10
0.553
0.583
Ranking quality at top 10 results
Recall@5
0.700
0.740
% of relevant docs in top 5
Recall@10
0.800
0.800
% of relevant docs in top 10
Latency Metrics
Mean
339ms
458ms
Average response time
P50
285ms
408ms
50th percentile (median)
P90
304ms
615ms
90th percentile

arguana

MetricCohere Rerank 3.5Cohere Rerank 4 ProDescription
Accuracy Metrics
nDCG@5
0.267
0.353
Ranking quality at top 5 results
nDCG@10
0.355
0.439
Ranking quality at top 10 results
Recall@5
0.520
0.660
% of relevant docs in top 5
Recall@10
0.800
0.920
% of relevant docs in top 10
Latency Metrics
Mean
570ms
785ms
Average response time
P50
373ms
768ms
50th percentile (median)
P90
617ms
933ms
90th percentile

FiQa

MetricCohere Rerank 3.5Cohere Rerank 4 ProDescription
Accuracy Metrics
nDCG@5
0.124
0.126
Ranking quality at top 5 results
nDCG@10
0.128
0.129
Ranking quality at top 10 results
Recall@5
0.123
0.130
% of relevant docs in top 5
Recall@10
0.130
0.135
% of relevant docs in top 10
Latency Metrics
Mean
364ms
610ms
Average response time
P50
315ms
585ms
50th percentile (median)
P90
401ms
817ms
90th percentile

business reports

MetricCohere Rerank 3.5Cohere Rerank 4 ProDescription
Latency Metrics
Mean
334ms
529ms
Average response time
P50
293ms
498ms
50th percentile (median)
P90
503ms
675ms
90th percentile

PG

MetricCohere Rerank 3.5Cohere Rerank 4 ProDescription
Latency Metrics
Mean
458ms
760ms
Average response time
P50
360ms
720ms
50th percentile (median)
P90
615ms
896ms
90th percentile

DBPedia

MetricCohere Rerank 3.5Cohere Rerank 4 ProDescription
Accuracy Metrics
nDCG@5
0.156
0.151
Ranking quality at top 5 results
nDCG@10
0.166
0.166
Ranking quality at top 10 results
Recall@5
0.004
0.004
% of relevant docs in top 5
Recall@10
0.005
0.005
% of relevant docs in top 10
Latency Metrics
Mean
286ms
541ms
Average response time
P50
279ms
489ms
50th percentile (median)
P90
290ms
729ms
90th percentile

Explore More

Compare more rerankers

See how all reranking models stack up. Compare Cohere, Jina AI, Voyage, ZeRank, and more. View comprehensive benchmarks, compare performance metrics, and find the perfect reranker for your RAG application.