Code search without a price hike: Voyage rerank-3 on OpenRouter costs the same as the previous generation

A reranking model takes a search query and documents already retrieved, then ranks them by relevance. In repository search, it selects which code snippets to show first.

Replacing earlier models. Voyage AI offers rerank-3 as a drop-in replacement for rerank-2.5, and rerank-3-lite for rerank-2.5-lite. The new generation is designed for long documents and code search.

OpenRouter measures quality using NDCG@10: how accurately the model ranks the first 10 results. Its published measurements from Sep 30, 2026 show the following gains over the corresponding generation 2.5 model:

| Test | rerank-3 | rerank-3-lite | | --- | --- | --- | | Average across domains | 0.80% | 0.94% | | Long documents | 3.35% | 1.86% | | Code search after voyage-3-large | 2.01% | — | | Code search after voyage-4-large | 1.96% | — |

According to OpenRouter, rerank-3-lite matches the earlier rerank-2.5 in search quality. In other words, the new generation's lightweight model has reached the level of the previous generation's larger model.

OpenRouter's prices as of Sep 30, 2026 remain unchanged from generation 2.5:

| Model | API ID | Price per 1 million tokens | | --- | --- | --- | | rerank-3 | `voyageai/rerank-3` | $0.05 | | rerank-3-lite | `voyageai/rerank-3-lite` | $0.02 |

Connecting through OpenRouter. These models require a separate API call, rather than a button in chat. Send the `Authorization: Bearer <API_KEY>` header and JSON containing the query and documents to `POST https://openrouter.ai/api/v1/rerank`:

```json { "model": "voyageai/rerank-3", "query": "Where are user permissions checked?", "documents": ["First code snippet", "Second code snippet"], "top_n": 2 } ```

The `top_n` field sets the number of documents returned. For the lightweight model, use `voyageai/rerank-3-lite` in the `model` field.

Connecting directly to Voyage AI starts with `pip install -U voyageai`, a key from `dashboard.voyageai.com`, and the `VOYAGE_API_KEY` environment variable. This Python call returns 3 results:

```python import voyageai

result = voyageai.Client().rerank( query, documents, model="rerank-3", top_k=3, ) ```

The direct REST integration uses `POST https://api.voyageai.com/v1/rerank`. Voyage AI's pricing as of Sep 30, 2026 lists 200 million free tokens for each new model. When calculating usage, the service counts the search query's tokens separately for each document.

Voyage AI's direct API limits for both models as of the same date:

- Up to 32,000 tokens per search query and document pair. - Up to 8,000 tokens in the search query. - Up to 1,000 documents per call. - Up to 600,000 counted tokens per call.

Both models are available through OpenRouter and directly through the Voyage AI API.