Learn  /  What's new

BigQuery's November 2026 update: Gemini 4 in SQL, project caps for slots, JSON_FLATTEN and federated queries to PostgreSQL, MySQL and Oracle

Google's big data cloud update grabbed the headlines at the start of the month. The smaller release notes that followed are the ones you will actually use on a Tuesday afternoon. Here they are, sorted by who they help.

What's new7 min readUpdated 16 November 2026
SQLDatabasesWhat's new
All guides and cheat sheets

Our article on Google's 4 November announcement covered BigQuery Graph and Knowledge Catalog. This one covers the release notes from 5 to 13 November 2026: smaller features, many of them generally available (GA) straight away. The October update covers the month before.

17Release notes5 to 13 November
11Generally availablethe rest in preview or rollout
2New Gemini modelsin generative AI functions
3New database connectionsMySQL, Oracle, PostgreSQL

For analysts

Gemini 4 in your SQL

BigQuery's AI functions, such as AI.CLASSIFY, AI.IF and AI.GENERATE, can now use gemini-4-carbon (10 November) and gemini-4-hydrogen (12 November). If you do not name a model, BigQuery picks one for the best balance of cost and quality, which for simple labelling is usually the right choice.

SQL
SELECT
  title,
  AI.CLASSIFY(
    body,
    categories => ['tech', 'sport', 'business', 'politics', 'entertainment', 'other']
  ) AS category
FROM `bigquery-public-data.bbc_news.fulltext`
LIMIT 100;

JSON_FLATTEN

Nested arrays in JSON are a common pain. JSON_FLATTEN (GA, 11 November) returns every non-array value found directly in the input or inside nested arrays, as an ARRAY<JSON>. These examples are from Google's documentation:

SQL
SELECT JSON_FLATTEN(JSON '[[[1]], 2, [3]]') AS json_flatten;
-- [1, 2, 3]

SELECT JSON_FLATTEN(JSON '[[[1, 2], 3], {"a": 4}, true]') AS json_flatten;
-- [1, 2, 3, {"a": 4}, true]

Note that objects are kept whole: JSON_FLATTEN(JSON '{"a": [[1]]}') returns [{"a":[[1]]}]. Combine it with UNNEST to get one row per value.

Graph, now on on-demand pricing

BigQuery Graph and GQL queries now work for projects on on-demand pricing (6 November), not only with reservations. Graph measures (GA, 10 November) lock an aggregation to a key, so totals do not overcount when a pattern matches the same node several times.

For admins and FinOps

Project caps (GA, 11 Nov)
Set the maximum slots and concurrency for one project inside a reservation, so a single heavy project cannot starve the others. Enterprise and Enterprise Plus editions.
Principal on reservation assignments (GA, 11 Nov)
Route queries to a specific reservation based on who runs them, a user or a service account. Useful for giving scheduled pipelines their own capacity.
Compute cost attribution (Preview, 9 Nov)
Opt in to charge compute costs directly to the project that ran the query, which makes chargeback much simpler.
Jobs explorer (GA, 9 Nov)
Troubleshoot a slow query and compare job and system performance side by side.
Pipeline task concurrency (GA, 9 Nov)
Set how many tasks of a pipeline may run at the same time.
Custom organization policy for models (Preview, 6 Nov)
Allow or deny specific operations on BigQuery ML models across the organization.

For data engineers

Generally available

Use now
  • Table-level data lineage in Knowledge Catalog for BigQuery pipelines (12 Nov)
  • Query text heatmap shows the SQL of your UDFs and a list of context references (5 Nov)

Lineage plus the heatmap make "why is this table wrong, and why is it slow?" much quicker to answer.

Preview

Try in a test project
  • Connections to MySQL, Oracle and PostgreSQL for federated queries (13 Nov)
  • Zstandard compression and decompression functions in GoogleSQL (9 Nov)
  • Conversational analytics in the Tokyo region (5 Nov)

Federated queries are handy for small lookups; copy large tables instead.

What to do this week

  1. 1

    Check your noisy projects

    If one project regularly eats your reservation, give it a project cap.

  2. 2

    Give pipelines their own capacity

    Use the principal setting to route a service account to its own reservation.

  3. 3

    Clean up JSON code

    Replace hand-written nested UNNESTs with JSON_FLATTEN where it fits.

  4. 4

    Test Hydrogen on one AI job

    Compare its labels and cost with your current model on a sample of 500 rows.

Check your understanding

1/3What do project caps limit?

2/3What does JSON_FLATTEN return for JSON '[[[1]], 2, [3]]'?

3/3What changed for BigQuery Graph on 6 November?

More: BigQuery vs Snowflake, SQL window functions explained and AI functions in SQL compared.

Sources: BigQuery release notes, entries from 5 to 13 November 2026; Google Cloud documentation for AI.CLASSIFY and JSON functions (examples). Preview features may change before general availability.

Alessandro Ecclesie AgazziWritten by Alessandro Ecclesie Agazzi, analytics engineer in London. Updated 16 November 2026. About me

Found this useful?

New guides and "what's new" explainers on data and AI land here regularly. Follow along on LinkedIn, or share this one with a colleague.