What is Google semantic-ranker-005?
semantic-ranker-default-005 and semantic-ranker-fast-005 are the models behind Google Cloud's Ranking API, released on 1 September 2026. The API takes a query and up to 1,000 records, each a title and some content, and returns them reordered with a relevance score between 0 and 1. It is stateless, so nothing has to be indexed first, and it sits on top of whatever search system produced the candidates.
Both models read text in 25 languages and look at up to 1,024 tokens per record; longer records are split and scored in segments. The @latest aliases now point to the 005 models, so callers that never pinned a version moved to them automatically. Google has published no benchmark or announcement for 005; the last public figures are for the 004 generation from 2025.
Pricing is per query rather than per token: $1.00 per 1,000 queries, where one query covers up to 100 records. It runs on Google Cloud only, through the Discovery Engine API, Vertex AI RAG Engine, AlloyDB and the Vector Search 2.0 reranking step.
How much does Google semantic-ranker-005 cost?
| Plan | Price | What you get |
|---|---|---|
| Pay as you go | $1.00 per 1,000 queries | Up to 100 records per query |
| 1-year savings plan | $0.90 per 1,000 queries | |
| 3-year savings plan | $0.80 per 1,000 queries |
$1.00 per 1,000 queries pay as you go, $0.90 on a 1-year and $0.80 on a 3-year savings plan. One query covers up to 100 records; 132 records count as 2 queries.
Is Google semantic-ranker-005 worth it?
The billing model decides whether this API makes sense for you. Google charges $1.00 per 1,000 queries no matter how long the records are, while Cohere and Voyage charge per token. A typical RAG call that reranks 20 passages of 300 tokens costs $0.001 here and about $0.0003 on Cohere Rerank 5 Pro at $0.05 per million tokens, so short shortlists are roughly three times dearer on Google. Send 100 records of 1,000 tokens instead and the order flips: still $0.001 here, about $0.005 on Cohere Pro.
What the 005 models change is not documented. The docs list them, give a release date of 1 September 2026 and point the @latest aliases at them, and nothing more: no launch post, no benchmark, no latency figure. The public evidence is for 004, which Google said in May 2025 led standalone reranking APIs on BEIR at NDCG@5, ran at least twice as fast as competing services, and had a fast variant with about a third of the default model's latency. Those claims are Google's own and a generation old.
The 1,024-token window per record is short next to the 32K pairs that Cohere and Voyage now accept. Longer records are chunked and scored in segments, and the docs warn that very long records may be partly ignored. For product listings, support articles and chunked RAG passages that rarely matters; for whole contracts it does.
The practical case is integration. On Google Cloud the ranker is already wired into Vertex AI RAG Engine, AlloyDB's ai.rank() SQL function and Vector Search 2.0, which gives each project 80,000 free ranking units every 30 days and falls back to its own fused results if reranking fails. Outside Google Cloud there is little reason to choose it over rerankers with longer context and published scores.
Where it earns its place
A search engineer on a retailer's product catalogue running on Google Cloud
Put the products that match a shopper's phrasing at the top of 100 keyword results
Send the top 100 product titles and descriptions to semantic-ranker-fast@latest; at one query per 100 records that is $1.00 per 1,000 searches, or $1,000 for a million, whatever the description length.
A data engineer whose application data lives in AlloyDB
Reorder vector-search results inside SQL without a separate service
Call ai.rank() on the candidate rows in the same query, so ranking happens next to the data. Google's docs describe the function; costs follow the Ranking API's per-query price.
A platform team on Vertex AI Vector Search 2.0
Add a reranking step after hybrid search without new infrastructure
Turn on the VertexRanker step after RRF fusion; each project gets 80,000 free ranking units every 30 days, and if reranking fails the call still returns the fused results with a warning.
Before you commit
- Setup
- Enable the Discovery Engine API on a Google Cloud project, then POST a query and records to the rankingConfigs rank endpoint. Leaving out the model uses semantic-ranker-default@latest, now the 005 model.
- Integrations
- REST and client libraries for the Discovery Engine API; Vertex AI RAG Engine, AlloyDB ai.rank(), Vector Search 2.0, LangChain, Genkit and Elasticsearch's Open Inference API.
- Data and privacy
- Google states that customer data used in Agent Search is not used to train foundation models; terms follow Google's Cloud Data Processing Addendum. Requests in the documented examples go to the global location.
- Cost at scale
- $1.00 per 1,000 queries, $0.90 or $0.80 with 1- or 3-year savings plans. Each block of 100 records counts as one query, so 2 million calls of 20 records cost about $2,000 a month.
- Limits
- 1,024 tokens per record (title plus content), 1,000 records per request and 500 requests per minute per project by default. Vector Search 2.0 still lists only the 004 versions and the @latest aliases.
What’s good
- Flat $1.00 per 1,000 queries, whatever the record length
- Up to 1,000 records per request; no indexing needed
- Built into RAG Engine, AlloyDB and Vector Search 2.0
- Customer data is not used to train Google's foundation models
What’s not
- No published benchmarks or announcement for the 005 models
- 1,024 tokens per record, against 32K on Cohere and Voyage
- Google Cloud only, through the Discovery Engine API
- Text only; no documented instruction input
Best for
Teams already building search or RAG on Google Cloud who rerank long candidate lists and want per-query billing.
Google semantic-ranker-005 alternatives
Google semantic-ranker-005: common questions
- Is Google semantic-ranker-005 worth it?
- Google's reranker is the cheap choice for long shortlists already on Google Cloud, but the 005 models arrived with no published numbers to judge them by. We rate it 3.9 out of 5.
- How much does Google semantic-ranker-005 cost?
- $1.00 per 1,000 queries pay as you go, $0.90 on a 1-year and $0.80 on a 3-year savings plan. One query covers up to 100 records; 132 records count as 2 queries.
- What are the downsides of Google semantic-ranker-005?
- No published benchmarks or announcement for the 005 models; 1,024 tokens per record, against 32K on Cohere and Voyage; google Cloud only, through the Discovery Engine API; text only; no documented instruction input.
- Who is Google semantic-ranker-005 best for?
- Teams already building search or RAG on Google Cloud who rerank long candidate lists and want per-query billing.
- How much does Google semantic-ranker-005 cost for a team or in production?
- $1.00 per 1,000 queries, $0.90 or $0.80 with 1- or 3-year savings plans. Each block of 100 records counts as one query, so 2 million calls of 20 records cost about $2,000 a month.
- Is Google semantic-ranker-005 safe for company data?
- Google states that customer data used in Agent Search is not used to train foundation models; terms follow Google's Cloud Data Processing Addendum. Requests in the documented examples go to the global location.
- What are the best Google semantic-ranker-005 alternatives?
- The closest alternatives we have reviewed are Cohere Rerank 5, Voyage rerank-3, zerank-2 and Jina Reranker v3.5.
Answers checked Nov 20, 2026.