AI Vector Search itself carries no separate license charge inside Enterprise Edition, Standard Edition 2, or Autonomous, but the way you deploy it can quietly pull in chargeable options and expose undocumented servers. This page pins down exactly where the feature sits in the 23ai model, what compute it consumes, and what actually triggers a bill at audit.
AI Vector Search itself carries no separate license charge inside Enterprise Edition, Standard Edition 2, or Autonomous, but the way you deploy it can quietly pull in chargeable options and expose undocumented servers. This page pins down exactly where the feature sits in the 23ai model, what compute it consumes, and what actually triggers a bill at audit.
Yes, AI Vector Search is included. Oracle ships it inside Oracle Database Enterprise Edition, Standard Edition 2, and Autonomous Database at no additional license charge. There is no line item called "Vector Search Option," no separately priced pack, and no re-certification hoop. Oracle's own AI Database FAQ states plainly that advanced AI features, including AI Vector Search, are included at no additional charge, with no database upgrade required. For an EE shop already paying $47,500 per processor, that is a genuine economic lever: workloads that would otherwise buy a standalone vector store (Pinecone, Weaviate, Milvus, Chroma, Qdrant) can run inside a database you already license and support.
That is the answer Oracle sales will give you, and it is accurate as far as it goes. The long answer, the one that matters at audit, is that the feature is free but the surrounding architecture frequently is not. Enabling Vector Search on the wrong server, or scaling it the way Oracle's documentation recommends, can pull in Partitioning, Real Application Clusters (RAC), Advanced Security, and Diagnostics Pack, every one of which carries a separate license. The trap is not the feature itself. It is what the feature invites you to deploy alongside it. Our broader treatment of this sits in the Oracle 23ai vector and AI feature licensing pillar, and this page drills into the specific EE-versus-extra question.
The feature is free. The architecture Oracle recommends around it is where the invoice hides.
Vector Search is a base-database capability, not an option. That distinction is the whole story. Options such as Partitioning ($11,500 per processor), RAC ($23,000 per processor), and the Diagnostics Pack are entitlements you buy on top of your edition and that Oracle's LMS scripts detect the moment you touch them. Base features carry no such flag. Vector Search behaves like SQL or PL/SQL: it is simply present.
Notably, Oracle extended Vector Search down into Standard Edition 2, which is unusual for a feature Oracle markets as flagship AI. Oracle's guidance confirms that many developer features, including AI Vector Search, are available in SE2, subject to the usual SE2 constraints (two-socket ceiling, no Partitioning, no RAC, no parallel query). Because EE lists at roughly 2.7 times SE2 per processor and per Named User Plus, an SE2-eligible vector workload that fits inside a two-socket box is a materially cheaper home. If your vector use case is a bounded corpus and modest concurrency, do not assume you need EE. Compare against our SE2 versus Enterprise Edition breakdown before defaulting to EE out of habit.
One more model detail worth banking: the license framework introduced with 23c carries forward unchanged into 23ai, and the October 2025 rebrand to "26ai" changes nothing in licensing. The banner will read 23.26.1, but the metrics, prices, and option boundaries are identical. Wherever a contract or renewal quote references 23c, treat it as applying to 26ai. Do not let a vendor rep use the rename to reopen pricing you have already settled.
Because Vector Search itself is $0 in license terms, the only numbers that matter are the database license and support underneath it. Those are unchanged in 23ai/26ai:
| Item | Enterprise Edition | Standard Edition 2 |
|---|---|---|
| License per Processor | $47,500 | approx. $17,500 (roughly 2.7x cheaper) |
| License per Named User Plus | $950 (25 NUP min per proc) | approx. $350 (10 NUP min per server) |
| Annual support | 22% of license, year over year | 22% of license, year over year |
| AI Vector Search | Included, $0 | Included, $0 |
| Partitioning (needed for IVF at scale) | $11,500 per Processor | Not available |
| RAC (needed for HNSW clustering) | $23,000 per Processor | Not available |
Work a concrete example. Ten EE processors is $475,000 in license plus $104,500 in first-year support. Add Partitioning across those ten processors and you have added $115,000 in license and $25,300 in annual support, before you have run a single similarity query. That is the gap between "the feature is free" and "the deployment is not." The database cost, not the vector feature, is what you are governing. Our Oracle Database licensing reference holds the full option and pack price list.
Licensing aside, Vector Search has a real compute and memory profile that shapes both performance and your option footprint. There are two index types with very different characteristics, and choosing between them has license consequences.
Hierarchical Navigable Small World (HNSW) indexes are in-memory neighbor graphs. Oracle stores them in a new SGA pool, the Vector Pool, sized by the VECTOR_MEMORY_SIZE parameter. The default is 0, and the ceiling is 70% of SGA. It is modifiable at the pluggable database level. The trap here is operational rather than contractual, but it bites hard: with VECTOR_MEMORY_SIZE set to 0, Oracle builds the HNSW graph but has nowhere to hold it, so the optimizer falls back to a full scan and performance collapses. Teams then over-provision hardware to compensate, and in Oracle's per-processor world, more cores means more license. The memory setting is free; the cores you buy to feed it are not.
Inverted File Flat (IVF) indexes are the disk-based alternative for datasets too large to hold in memory. They cluster vectors into partitions by similarity and support both global and local partitioning on partitioned tables. Read that carefully: the recommended path for large IVF deployments involves partitioned tables, and Partitioning is a separately licensed EE option at $11,500 per processor. If your DBAs build IVF indexes on partitioned base tables, or use partitioning to manage the vector data lifecycle, you have crossed into chargeable territory even though the vector feature was free. This is the single most common way a "free" AI feature turns into a Partitioning finding.
Vectors are storage-heavy. At 1,536 dimensions in FLOAT32, each vector consumes roughly 6 KB. Ten million embeddings is about 60 GB of vector data before relational columns. That drives SGA sizing (for HNSW), storage cost, and, if you resort to more powerful hardware, core counts. Note that no GPU is required: Oracle confirms vectors can be created on CPUs or GPUs. Do not let a vendor or integrator upsell GPU-heavy engineered systems as a Vector Search prerequisite. It is not one.
VECTOR_MEMORY_SIZE is free. The extra cores teams buy to compensate for setting it wrong are $47,500 each.
This is the section that earns its place. Oracle's own GA announcement states that AI Vector Search leverages Partitioning, RAC, Sharding, and Exadata for scalability; Data Guard, GoldenGate, Flashback, RMAN, and ZDLRA for availability; and Advanced Security (TDE, Key Vault, Audit Vault, VPD) for security. Every one of those is separately licensed. Oracle markets them as things Vector Search "leverages." In license terms, every one of them is a potential invoice. Four traps stand out.
The pattern is consistent: the feature is free, the reference architecture is not. Before any team enables Vector Search in production, run an option-usage review against the specific hosts. Our note on the license review to run before you migrate to 23ai lays out the mechanics, and it applies whether you are upgrading or just switching on vector features on an existing 23ai instance.
There is genuinely good news to bank at your next renewal, and it has nothing to do with Vector Search directly. Oracle folded several previously separately licensed options into the EE base in 23ai, with Spatial and Graph the most material. If you were paying for those options on 21c or 19c, upgrading to 23ai removes a legitimate cost, and that is a negotiation point at your support renewal or true-up. Do not let it be quietly absorbed. Quantify what Spatial and Graph were costing you and demand the reduction. This is the kind of offset you use to fund the Partitioning or Advanced Security you genuinely need for a serious vector deployment, so the net stays flat rather than climbing.
There is one comparison worth running in parallel: whether Autonomous Database changes the math, since AI features are included there too but the compute metric differs. We treat that trade-off in running 23ai AI features on Autonomous versus self-managed. For heavily variable vector workloads, the Autonomous consumption model sometimes beats fixed processor licensing, and sometimes does not.
Oracle announced 23ai in May 2024, but broad on-premises availability for the AI Database (rebranded 26ai) arrived only in 2026, having previously been confined to engineered systems and cloud. This matters for two reasons. First, if a rep claims you have been entitled to Vector Search on standard on-prem hardware since 2024, check the ship dates for your platform; you may not have been. Second, moving from 23ai to 26ai is described as simply applying the October 2025 release update, with no license change. Confirm in writing that the 26ai banner does not trigger any re-quoting. Get the "no licensing change" statement from Oracle in writing, on the order or in an email, rather than relying on marketing FAQs that Oracle can revise.
Treat Vector Search as free and treat everything it touches as a compliance decision. Concretely: (1) confirm your target hosts are fully licensed for EE or SE2 before enabling the feature, because undocumented databases remain the top finding; (2) choose HNSW versus IVF with the option consequences in front of you, since IVF at scale pulls in Partitioning and HNSW on a cluster surfaces RAC; (3) set VECTOR_MEMORY_SIZE deliberately so you are not tempted to buy cores to fix a parameter mistake; (4) claim the Spatial and Graph consolidation as a renewal offset; and (5) get Oracle's no-additional-charge and no-licensing-change-on-26ai positions in writing. If you are weighing SE2 as the cheaper home, or Autonomous as the variable-cost home, read the linked comparisons before committing. The feature costs nothing. The discipline around it is what keeps it that way.
No. It is included in Enterprise Edition, Standard Edition 2, and Autonomous Database at no additional license charge. It is a base-database feature, not an option like Partitioning or RAC, so there is no separate line item to buy. You still need a valid EE or SE2 database license underneath it.
Yes. Oracle confirms AI Vector Search is available in SE2, subject to the usual SE2 limits (two-socket ceiling, no Partitioning, no RAC, no parallel query). Since EE lists at roughly 2.7 times SE2, an SE2-eligible vector workload that fits inside a two-socket server is materially cheaper. Verify your corpus size and concurrency fit within SE2 constraints first.
IVF indexes at scale lean on Partitioning ($11,500 per processor), HNSW indexes on a RAC cluster surface RAC ($23,000 per processor) through full index duplication, and encrypting embeddings can pull in Advanced Security. None are included in EE. Run an option-usage review on the specific hosts before enabling the feature in production.
No. Oracle's 26ai is functionally 23ai with the October 2025 release update applied, and it appears as 23.26.1 in the banner. The metrics, prices, and option boundaries are unchanged. Get Oracle's no-licensing-change statement in writing so a rep cannot use the rename to reopen pricing.
HNSW indexes live in an SGA Vector Pool sized by VECTOR_MEMORY_SIZE (default 0, ceiling 70% of SGA). At 1,536 dimensions in FLOAT32, each vector is about 6 KB, so ten million embeddings is roughly 60 GB. No GPU is required; vectors can be created on CPUs or GPUs. Setting the memory pool to 0 forces full scans, so size it deliberately.
Yes, indirectly. Oracle folded several previously separate options into the EE base in 23ai, with Spatial and Graph the most material. If you were paying for those on 19c or 21c, upgrading removes that cost, and you should claim it as a renewal offset rather than letting Oracle absorb it silently.
Oracle SE2 licenses per occupied socket at 17,500 dollars, two sockets maximum, 16 thread cap. Free paper on the metric choice and Enterprise Edition trap.
Gated with a work email on the download page. No sales follow up you did not ask for.
Get the White Paper →
500+ enterprise clients. 11 vendor practices. Industry recognized. One conversation can change what you pay for the next three years.
One buyer side briefing a week. Renewal signals, audit moves, and the levers that work. No vendor spin.
