If your BigQuery estate was designed in 2021 or 2022 and has mostly been left alone since, it is running on a product that no longer exists in the form you built for. The SQL still works — Google is careful about that — but the pricing model, the tooling around the warehouse, the AI features, and the open-format story have all been replaced. This post is the catch-up briefing we give new clients, and it doubles as the announcement that we have refreshed this site to match: new AI and Gemini, cost optimization and lakehouse services, and a tutorials section that is finally full.
1. Pricing: flat-rate is gone, Editions are here
On July 5, 2023, Google retired flat-rate and flex slots. Compute is now either on-demand (per TiB scanned) or one of three Editions — Standard, Enterprise, Enterprise Plus — billed per slot-hour with an autoscaler that grows capacity in 50-slot steps between a baseline and a cap. Storage gained a second model too: physical billing, which is cheaper for compressible data and more expensive for datasets with long time-travel windows.
If you are still on 2022 patterns: you probably have a reservation sized for your peak hour and a storage bill nobody has examined. Both are fixable in a week; see our Editions explainer.
2. Gemini is in the product, not beside it
BigQuery Studio replaced the old console editor. Inside it, Gemini generates and explains SQL, data canvas lets analysts explore visually, and data preparation builds cleansing steps from suggestions. More consequentially, generative AI became SQL functions: AI.GENERATE, ML.GENERATE_TEXT, ML.GENERATE_EMBEDDING call Gemini over every row of a table. Vector indexes and VECTOR_SEARCH make BigQuery a credible retrieval layer for RAG. And in April 2026 the Data Engineering Agent went GA — natural-language pipeline generation into Dataform, reviewable as code.
If you are still on 2022 patterns: you are likely exporting data to a separate vector store and a separate LLM pipeline, paying twice for governance you could have inherited for free. Our vector search tutorial shows the in-warehouse alternative.
3. Tools were renamed, retired, or replaced
A short list that dates a codebase instantly:
- Stackdriver became Cloud Monitoring and Cloud Logging (2020)
- Data Studio became Looker Studio (2022)
- Cloud Dataprep is an Alteryx product; Google's answer is BigQuery data preparation plus Dataform
- Cloud Functions became Cloud Run functions (2024)
- Data Catalog was folded into Dataplex Universal Catalog
- Scheduled queries are still supported, but Dataform is where new pipeline work belongs
None of these break anything. All of them are signs that nobody has reviewed the estate recently.
4. Iceberg and the open lakehouse
BigQuery now manages Apache Iceberg tables natively — Parquet in your bucket, Iceberg metadata external engines can read, BigQuery handling DML, streaming, and storage optimisation. The BigLake metastore exposes an Iceberg REST catalog so Spark, Trino, and friends read the same tables without copies. The warehouse-versus-lake argument is over; the question is now which tables are native and which are Iceberg.
If you are still on 2022 patterns: you may be maintaining a Parquet lake and a BigQuery copy of it. One of them can go.
5. SQL itself moved
Pipe syntax (|>) is GA and makes multi-step analytics queries linear instead of nested. Search indexes mean "BigQuery has no indexes" is no longer true. Multi-statement transactions, JSON as a first-class type, and INFORMATION_SCHEMA views for nearly everything round out a dialect that is noticeably more capable than the one your 2022 queries were written in.
The 2026 migration checklist
- Re-benchmark on-demand versus Editions with real
INFORMATION_SCHEMA.JOBShistory; right-size or drop the reservation - Evaluate physical storage billing per dataset; shorten time-travel where seven days is overkill
- Turn on
require_partition_filterfor every large partitioned table - Replace retired product references in code, docs, and alerts
- Move new pipeline work to Dataform; pilot the Data Engineering Agent with a review gate
- Decide, table by table, what should be Iceberg
- Pick one AI use case with measurable value — ticket classification, document search — and build it inside the warehouse
- Put a FinOps dashboard in front of someone who will look at it weekly
We run this as a two-week assessment. If you would like one, get in touch.