Published: 2026-08-27

What's New in Oracle 26ai (and Where 23ai Went)

Oracle AI Database 26ai was announced at Oracle AI World on October 14, 2025 — and the first thing to understand is that it is not a new major release. 26ai replaces 23ai: applying the October 2025 Release Update turns a 23ai system into 26ai, with no upgrade and no application re-certification, and the versions 24 and 25 simply never existed. That makes "what's new" a genuinely confusing question, because most of what gets listed under 26ai — AI Vector Search, JSON Relational Duality, True Cache — shipped with 23ai back in May 2024. Below is the actual split: what the 23.26.x Release Updates added (SQL assertions, QUALIFY, Select AI everywhere, Deep Data Security, post-quantum crypto, DATEDIFF), what people conflate with it, the support clocks that decide your upgrade urgency, and — first-hand — what a desktop SQL client needs to do about any of it. Everything here is checked against Oracle's documentation and release notes as of August 2026.

The short version

Where 23ai went: the rename, the numbering, the dates

Oracle renamed the product family from "Oracle Database" to "Oracle AI Database" and jumped the marketing version from 23 to 26 so it tracks the calendar year — the same move MySQL made with its calendar-versioned 26.x line. Under the hood nothing jumped at all: the internal version stays 23, and the Release Update numbering now encodes year and quarter. Oracle's own upgrade PM spells it out: 23.26.0 is the October 2025 RU, 23.26.1 is January 2026, 23.26.2 April 2026, 23.26.3 July 2026. The 23ai-named RU train ended at 23.9 in July 2025.

The dates that matter:

• October 14, 2025 — 26ai announced at Oracle AI World; existing 23ai systems become 26ai by applying RU 23.26.0. Autonomous Database converted automatically.
• January 27, 2026 — on-premises GA for Linux x86-64 (23.26.1). Enterprise Edition only — the installer refuses Standard Edition 2 — with more platforms planned through 2026.
• July 7, 2026 — Amazon RDS for Oracle adds 26ai, as an upgrade target for 19c/21c CDBs and for new instances.

SQL assertions: constraints that can finally see the whole table

The most interesting pure-SQL feature in 26ai is CREATE ASSERTION (RU 23.26.1). A CHECK constraint can only look at the row being written; an assertion is a database-level CREATE ASSERTION … CHECK (…) whose condition can reference whole tables, aggregates, and joins across several tables — "every account's withdrawals must not exceed its balance", "an agent may have at most 25 open tickets". The engine enforces it on every commit that could violate it and raises ORA-08601 when one does. Assertions have been in the SQL standard since SQL-92; Oracle is the first of the major engines to actually ship them, which is why the feature is worth knowing about even if you never write one — it moves a whole class of trigger-and-application-code invariants into the schema.

QUALIFY: filtering window functions without the subquery dance

Snowflake, BigQuery, and DuckDB users have had this for years; Oracle added it in RU 23.26.0. QUALIFY is to analytic functions what HAVING is to GROUP BY — the classic "latest row per customer" stops needing an inline view:

-- Oracle 26ai: latest order per customer, no subquery
SELECT o.customer_id,
       o.order_id,
       ROW_NUMBER() OVER (PARTITION BY o.customer_id
                          ORDER BY o.placed_at DESC) AS rn
FROM   orders o
QUALIFY rn = 1;

Select AI and the AI plumbing

Natural-language-to-SQL profiles (Select AI) were an Autonomous Database exclusive; RU 23.26.1 puts them in every 26ai deployment, and on RDS they wire to Amazon Bedrock. Around it, 26ai builds out the AI infrastructure: a Private AI Services Container (RU 23.26.2–3) runs embedding models, LLMs, and rerankers next to the database and can offload vector-index builds to a GPU box, and the launch messaging leans hard on MCP server support for agentic access. Whether you want your database speaking to LLMs is a separate question — but if you do, the pieces are now in the box rather than in a cloud-only tier.

Security: Deep Data Security and post-quantum crypto

Deep Data Security ("Deep Sec", RU 23.26.2–3) moves row-, column-, and cell-level visibility rules into the engine, evaluated with the end user's identity — drivers grew a setEndUserSecurityContext() call so a connection-pooling app can tell the database who is actually asking. Post-quantum cryptography (RU 23.26.1) adds quantum-resistant algorithms to network encryption. Both are Enterprise-Edition territory, but they are the kind of feature that shows up in compliance questionnaires long before your own code touches them.

Vectors: 23ai invented them, 26ai makes the indexes production-grade

To be clear about the split: the VECTOR type, VECTOR_DISTANCE, and HNSW/IVF vector indexes are 23ai features from May 2024. What the 23.26.x RUs add is index maturity — scalar-quantized HNSW (smaller memory footprint), distributed HNSW on RAC, local HNSW on partitioned tables, online index rebuild, automatic IVF reorganization — plus utility functions like VECTOR_ORIGIN and VECTOR_RANDOM. The basics, which work on any 23ai-family server, look like this (syntax reference):

CREATE TABLE docs (
  id        NUMBER GENERATED ALWAYS AS IDENTITY PRIMARY KEY,
  body      CLOB,
  embedding VECTOR(768, FLOAT32)          -- formats: INT8, FLOAT32, FLOAT64, BINARY
);

SELECT id
FROM   docs
ORDER  BY VECTOR_DISTANCE(embedding, :query_vec, COSINE)
FETCH  FIRST 10 ROWS ONLY;

CREATE VECTOR INDEX docs_hnsw ON docs (embedding)
  ORGANIZATION INMEMORY NEIGHBOR GRAPH
  DISTANCE COSINE
  WITH TARGET ACCURACY 90;

Dimensions go up to 65,535 (BINARY vectors: up to 65,528, in multiples of 8), vector columns can't carry primary keys or appear in ORDER BY/GROUP BY/DISTINCT, and IVF indexes can't be built on SPARSE vectors while HNSW can.

The everyday-SQL batch

The features most working developers will actually touch first are the small ones:

• DATEDIFF and DATEADD (RU 23.26.1 / 23.26.3) — SQL Server-style date arithmetic, ending decades of MONTHS_BETWEEN-and-arithmetic idioms for cross-engine developers.
• Aggregate FILTER (23.26.1) — the SQL-standard COUNT(*) FILTER (WHERE …), long available in PostgreSQL.
• JOIN TO ONE (23.26.2) — declares a join as many-to-one so the optimizer can validate or eliminate it.
• Nested and correlated WITH clauses, WAIT n/NOWAIT on all DML (both 23.26.2), and partition keys as expressions (23.26.1).
• JSON index ergonomics (23.26.1–2) — dot-notation CREATE JSON INDEX, ALTER INDEX on JSON search indexes, DECODE/CASE inside SQL/JSON paths.
• And the genuine oddball: RU 23.26.3 teaches the database to speak the Redis wire protocol directly.

Already in 23ai — don't credit 26ai for these

If you're coming from 19c, the biggest news in 26ai is everything below — but it all shipped with 23ai in May 2024 (or in a 23ai-era RU), and any "26ai features" list that leads with these is recycling the 2024 launch:

• AI Vector Search and the VECTOR type (23ai base; BINARY format 23.5, SPARSE 23.6+).
• JSON Relational Duality views, SQL Firewall, True Cache, lock-free reservations, JavaScript (MLE) stored procedures, SQL property graphs.
• BOOLEAN columns, IF [NOT] EXISTS DDL, the VALUES table constructor, SELECT without FROM, domains and annotations.
• 4096-column tables (MAX_COLUMNS=EXTENDED), 1024-byte passwords, TLS 1.3, readable TNS error messages.
• From the late 23ai-era RUs (so pre-rename, mid-2025): GROUP BY ALL, SQL UUID generation, non-positional INSERT, time bucketing, sessionless transactions, JSON collections.

Support clocks and upgrade paths

Per Oracle's upgrade guide, only 19c and 21c upgrade directly to 26ai; older releases need a two-step path. The 19c dates were extended in February 2025, which took most of the panic out of that migration — the release with the real deadline is 21c, which loses all support in mid-2027. There is also a free edition for trying any of this: Oracle AI Database 26ai Free (2 CPUs, 2 GB RAM, 12 GB user data, container image available).

Connecting to 26ai from a desktop client — first-hand

Jam SQL Studio is a cross-platform desktop client (macOS, Windows, Linux) for Oracle, SQL Server, PostgreSQL, MySQL, and SQLite. Here is what 26ai actually required from us, checked against the driver we ship:

• Connectivity needed nothing. Because 26ai keeps internal version 23, the node-oracledb thin driver connects unchanged — thin mode supports server 12.1+, and the current driver line is tested against Oracle Database 26. Jam connects with no Oracle Instant Client, including on Apple Silicon — the setup (or rather, absence of one) is covered in our Instant-Client-free Oracle post.
• BOOLEAN results render as true/false — the driver maps the 23ai BOOLEAN type to a real JavaScript boolean (this is where older tools show 1/0).
• VECTOR values arrive as typed arrays (Float32Array, Int8Array, and friends), not strings — the driver we bundle (6.10) covers dense vectors of every format plus SPARSE, so a SELECT embedding FROM docs returns data rather than an error, and our Oracle adapter maps the type in the result metadata so the column reports as VECTOR.
• Autocomplete knows the new surface. As of Jam SQL Studio 1.4.24, the Oracle SQL autocomplete suggests QUALIFY, VECTOR_DISTANCE / TO_VECTOR / VECTOR_SERIALIZE / VECTOR_EMBEDDING, DATEDIFF / DATEADD, and CREATE ASSERTION, each tagged with the Oracle version that accepts it — a 19c server will reject QUALIFY, and the hint says so.

And the honest gaps, so nobody reads coverage into this that isn't there: Jam's visual Table Designer currently builds tables for SQL Server and PostgreSQL, not Oracle, so there's no GUI path to a VECTOR(768, FLOAT32) column definition — you write that DDL in the query editor (where the PL/SQL debugger and Package Explorer work against any 23ai-family server, 26ai included). There's no vector-similarity UI and no CREATE VECTOR INDEX dialog either. As of August 2026 that matches the rest of the desktop-client field — but we'd rather list the gap than imply the feature.

FAQ

When was Oracle 26ai released?

Oracle AI Database 26ai was announced on October 14, 2025 at Oracle AI World. It reached existing 23ai systems as the October 2025 Release Update (23.26.0), on-premises general availability came on January 27, 2026 for Linux x86-64 (23.26.1, Enterprise Edition only at first), and Amazon RDS for Oracle added 26ai on July 7, 2026. Other on-premises platforms are planned through 2026.

What is the difference between Oracle 23ai and 26ai?

They are the same database. The internal version number stays 23, and 26ai replaced 23ai via the October 2025 Release Update — Oracle states there is no database upgrade and no application re-certification. Oracle also skipped 24 and 25 entirely so the name matches calendar year 2026. What is genuinely new in 26ai is the content of the 23.26.x Release Updates: SQL assertions, the QUALIFY clause, Select AI in every deployment, Deep Data Security, post-quantum cryptography, vector index upgrades, and DATEADD/DATEDIFF. Headline features like AI Vector Search, JSON Relational Duality views, and BOOLEAN are 23ai features from 2024.

Is Oracle 23ai still supported?

Yes — by becoming 26ai. The 23ai-named Release Update train ended at 23.9 (July 2025); staying patched means applying the 23.26.x updates, which is also what turns a 23ai system into 26ai. The Lifetime Support Policy carries over: Premier Support for 26ai runs to December 2031, with Extended Support dates still to be announced.

Is Oracle 26ai free?

The full product is paid (on-premises currently Enterprise Edition only), but Oracle AI Database 26ai Free exists with the usual Free-edition limits — up to 2 CPUs, 2 GB of RAM, and 12 GB of user data — distributed as an RPM, a Windows build, and a container image. Oracle also includes AI Vector Search in the paid editions at no additional charge.

Can I upgrade directly from Oracle 19c to 26ai?

Yes. Direct upgrade to 26ai is supported from 19c and 21c only; anything older needs a two-step path. 23ai systems don't upgrade at all — they just apply the 23.26.x Release Update. The clocks: 19c Premier Support runs to December 31, 2029 (Extended to December 31, 2032), while 21c Premier Support ends July 31, 2027 with no Extended Support — so 21c users have the harder deadline.

Do existing drivers and SQL clients connect to Oracle 26ai?

Yes. Because 26ai keeps internal version 23, existing thin-mode drivers connect unchanged — node-oracledb's thin mode supports server 12.1 and later, and the current driver line is tested against Oracle Database 26. Driver upgrades only matter for type coverage: node-oracledb 6.5+ maps VECTOR columns, 6.8+ adds SPARSE vectors, and 7.0 adds 26ai-specific extras like pipelining and the Deep Data Security end-user context. Jam SQL Studio connects to 26ai in thin mode with no Oracle Instant Client, including on Apple Silicon.