## Saturday, December 13, 2025 # Cohere Rerank 4: A real upgrade over 3.5 Umida Muratbekova RAGAIReranker Cohere released Rerank 4 in two variants – [Rerank 4 Pro](/content/rerankers/cohere-rerank-4-pro/index.html) and [Rerank 4 Fast](/content/rerankers/cohere-rerank-4-fast/index.html) – so we plugged both into our benchmark to see: did they actually improve over v3.5, and how do they compare to each other? The results weren’t small bumps. Both are clear upgrades, but with different trade-offs. ## Cohere went from mid-pack to #2 Cohere’s previous reranker (v3.5) sat in the lower half of the leaderboard, with an ELO around 1457. - **[Rerank 4 Pro](/content/rerankers/cohere-rerank-4-pro/index.html)** reaches 1627 ELO (#2 overall) just below [zerank-2](/content/rerankers/zerank-2/index.html). For a single version upgrade, that’s a big shift. - **[Rerank 4 Fast](/content/rerankers/cohere-rerank-4-fast/index.html)** lands at 1506 ELO (#7) – still a meaningful improvement over v3.5 (+54 ELO), but clearly below Pro. So Rerank 4 isn’t one model; it’s two different points on the quality–latency curve. ## Patterns That Explain The Shift - **Biggest gains on “enterprise” workloads**: the largest improvements show up on business- and finance-style tasks, where [Rerank 4 Pro](/content/rerankers/cohere-rerank-4-pro/index.html) gains a bit over +400 ELO vs v3.5, and [Rerank 4 Fast](/content/rerankers/cohere-rerank-4-fast/index.html) gains around +300 ELO (business) and +140 ELO (finance). - **Pro has no regressions; Fast is more mixed**: [Rerank 4 Pro](/content/rerankers/cohere-rerank-4-pro/index.html) beats v3.5 on all 6 datasets, with every bar in the improvement plot above zero. [Rerank 4 Fast](/content/rerankers/cohere-rerank-4-fast/index.html) improves on enterprise and entity-heavy workloads, but gives up quality on some argumentation and web-style QA datasets. - **Latency is the main trade-off**: [Rerank 4 Pro](/content/rerankers/cohere-rerank-4-pro/index.html) is ~60% slower than [zerank-2](/content/rerankers/zerank-2/index.html) and sits around the same speed as Voyage 2.5. [Rerank 4 Fast](/content/rerankers/cohere-rerank-4-fast/index.html) is noticeably quicker than Pro, and both variants stay under a second on average. Let’s look at each finding in detail. ## The big gains are on business and finance-style content The overall ELO jump is nice, but the more interesting part is _where_ the gains come from. When we split the results by workload type, the biggest improvements show up on: - long, dense business documents - finance and economics question answering On those tasks, **[Rerank 4 Pro](/content/rerankers/cohere-rerank-4-pro/index.html)** scores roughly **one-third higher** than v3.5 in ELO terms. That’s a large gap for models that are running on identical input. **[Rerank 4 Fast](/content/rerankers/cohere-rerank-4-fast/index.html)** also improves on these workloads. The gains are smaller than Pro’s, but still clearly ahead of v3.5, especially on business reports and financial Q&A. This lines up well with how enterprise RAG systems actually look in practice: long internal reports, slide decks, memos, filings, and specialised vocabulary. In that environment, getting the “right” two or three passages into the context window matters more than shaving off a couple of milliseconds. So if your corpus leans towards board reports, financial notes, and internal strategy docs instead of short web snippets, [Rerank 4 Pro](/content/rerankers/cohere-rerank-4-pro/index.html) is not just marginally better – it’s operating in a different range than v3.5, with [Rerank 4 Fast](/content/rerankers/cohere-rerank-4-fast/index.html) sitting in between the two. ## Pro is consistent, Fast makes trade-offs Another thing we checked for was regressions. A common pattern with new models is: big gains on one dataset, small but real drops on another. That forces you into routing logic or model mixing to avoid the weak spots. **[Rerank 4 Pro](/content/rerankers/cohere-rerank-4-pro/index.html)** doesn’t show that pattern. Across all six workloads we test – argumentation, business reports, finance, long-form narrative, web-style QA, and more entity-heavy queries – Pro is ahead of v3.5. The size of the gain varies: - on enterprise-style tasks, the jump is large - on entity-focused workloads, the improvement is modest but still positive In practice, if you’re already using v3.5, you can swap it for **[Rerank 4 Pro](/content/rerankers/cohere-rerank-4-pro/index.html)** and expect better ranking quality across all the workloads we tested. For **[Rerank 4 Fast](/content/rerankers/cohere-rerank-4-fast/index.html)**, the picture is different: it improves on business, finance, and entity-heavy queries, but does worse than v3.5 on argumentation-heavy and web-style QA datasets. That makes Fast a good fit for enterprise-heavy workloads, but less of a general-purpose upgrade than Pro. ## You lose ~27% of speed for better quality The one real trade-off we see is latency. **[Rerank 4 Pro](/content/rerankers/cohere-rerank-4-pro/index.html)** stays under one second per request in our setup, but it is slower than the fastest model on the leaderboard: - roughly 60% slower than [zerank-2](/content/rerankers/zerank-2/index.html) - about the same speed as Voyage’s premium reranker **[Rerank 4 Fast](/content/rerankers/cohere-rerank-4-fast/index.html)** is noticeably quicker than Pro – its latency is about 37% lower (447 ms vs ~614 ms), which works out to roughly a 27% gain in throughput if everything else stays the same. In exchange, it gives up a meaningful amount of ELO compared to Pro. So the picture is: - if you want the absolute best speed–quality balance, [zerank-2](/content/rerankers/zerank-2/index.html) is still in front - if sub-second latency is fine and you care more about ranking quality on messy enterprise text, [Rerank 4 Pro](/content/rerankers/cohere-rerank-4-pro/index.html) is a strong choice - if you’re latency-sensitive and your data is mostly business and finance documents, [Rerank 4 Fast](/content/rerankers/cohere-rerank-4-fast/index.html) can be a reasonable compromise. ## Conclusion Taken together, the results are straightforward: **Rerank 4 is a real upgrade for Cohere**. - **[Rerank 4 Pro](/content/rerankers/cohere-rerank-4-pro/index.html)** moves from middle of the pack to reliably near the top, with the biggest gains on business and finance-style text and no regressions across the workloads we tested. - **[Rerank 4 Fast](/content/rerankers/cohere-rerank-4-fast/index.html)** gives a smaller but still clear improvement over v3.5, with better latency, and works best when your data is mostly enterprise content rather than mixed web-style queries. You do pay more in latency for Pro, but the quality jump is clear. Fast sits in the middle: better than v3.5, faster than Pro, but with more trade-offs. If you want to look closer, you can check our [reranker leaderboard](/content/rerankers/index.html) and compare [Rerank 4 Pro](/content/rerankers/cohere-rerank-4-pro/index.html) and [Rerank 4 Fast](/content/rerankers/cohere-rerank-4-fast/index.html) against the other models we’ve benchmarked side by side.