Oracle ships AI Vector Search inside Enterprise Edition at no added license cost, then surrounds it with options that switch on through ordinary use. This guide separates what is genuinely free from what will show up in your next audit finding.
Oracle ships AI Vector Search inside Enterprise Edition at no added license cost, then surrounds it with options that switch on through ordinary use. This guide separates what is genuinely free from what will show up in your next audit finding.
Oracle Database 23ai (rebranded in later documentation as Oracle AI Database 26ai) is the first release Oracle has marketed on artificial intelligence rather than on core database mechanics. The headline message is generous: AI Vector Search is included in Enterprise Edition at no additional license cost, and it even ships in the free edition. That statement is true. It is also the most dangerous sentence in the release, because it conditions buyers to assume the whole AI stack is free. It is not.
In 25 years of negotiating Oracle Database agreements from the buyer side, the pattern is consistent: Oracle gives away the feature everyone talks about and charges for the features you need to make it useful at scale. The vector data type is free. The Real Application Clusters and Partitioning you need to scale vectors across instances are not. The embedding models load through PL/SQL packages that also touch machine learning territory. And every option you touch, even once, is recorded permanently in a view that Oracle License Management Services (LMS) examines as standard audit practice.
This article maps the exact boundary. It tells you which 23ai capabilities are genuinely included in Enterprise Edition (EE), which quietly require separately licensed options or management packs, and where 23ai adoption creates fresh audit exposure that buyers who upgraded from 19c do not see coming. Where a claim is contested between sources, we flag the conflict rather than paper over it, because those conflicts are exactly the points Oracle will exploit in an audit and exactly the points you must nail down in contract language before you deploy.
Start with the good news, because it is real and you should use it as leverage. AI Vector Search adds a native VECTOR data type, vector embedding storage, two index types (HNSW and Inverted File Flat), and both approximate and exact nearest-neighbor search directly into the database engine. Oracle ships all of this in Enterprise Edition at no additional license cost. Oracle's public position, stated at launch, is that AI functionality should be included in the price of the database rather than sold as an extra-cost option. That position is documented and you should hold Oracle to it in writing.
The inclusion runs deeper than EE. AI Vector Search functionality appears in Oracle Database Free and in the Autonomous Database Free container image. Developers can build semantic search, retrieval-augmented generation (RAG) pipelines, and vector database prototypes on the free edition with no license at all. That matters for two reasons. First, it kills any vendor argument that vector search is a premium capability you must pay to unlock. Second, it means your development and test teams can validate 23ai AI workloads without generating any license liability, provided they stay on Free edition and off production hardware.
True Cache is also included in Enterprise Edition at no additional cost. True Cache is a mostly diskless, read-only, automatically managed replica of the primary database, useful for offloading read traffic. Oracle's own True Cache FAQ confirms the feature is bundled with the EE license. JSON Relational Duality Views, which let the same data be accessed as both relational rows and JSON documents, function identically across editions and carry no separate charge. Oracle's converged database features (Property Graph, JSON, XML) are stated to be present in all Oracle Database offerings.
The feature everyone talks about is free. The infrastructure you need to run it at production scale is where Oracle earns its money.
So the free tier is genuinely broad: vector data type, vector indexing, semantic search, RAG primitives, True Cache, JSON Relational Duality, Property Graph, XML. If your 23ai use case fits entirely inside single-instance Enterprise Edition with in-memory HNSW indexes and no advanced security or management packs, you can adopt the AI features without a single incremental license. Most organizations do not stay inside that box for long, which is where the rest of this guide comes in. For a focused breakdown of the vector question specifically, see our analysis of whether Oracle 23ai AI Vector Search is included in Enterprise Edition or an extra.
Before you assume the free vector story applies to your estate, check your edition. Standard Edition 2 (SE2) cannot run AI Vector Search, full stop. SE2 also cannot run Real Application Clusters, Partitioning, Advanced Compression, Advanced Security, Active Data Guard, or Multitenant beyond three pluggable databases. SE2 remains capped at two sockets per server and a 16-CPU-thread runtime limit. Those limits are unchanged in 23ai.
This creates a specific 23ai pressure. Oracle's sales motion around AI features works only on Enterprise Edition, so any SE2 customer who wants native vector search is being steered toward an EE upgrade. Enterprise Edition costs roughly 2.7 times SE2 per Processor and roughly 2.7 times per Named User Plus (NUP). The 23ai release does not change the structural EE-versus-SE2 decision, but the inclusion of AI Vector Search in EE nudges the value calculation slightly toward EE for buyers who genuinely need vector workloads. The word to watch is 'genuinely'. Do not upgrade an entire SE2 estate to EE because one application team wants to experiment with RAG. Isolate the AI workload, license only that footprint on EE, and keep the rest on SE2.
For EE list pricing, published reference points put a Processor license at roughly 47,500 pounds sterling, and the US Processor list has historically sat near 47,500 US dollars before discount. Whatever your currency, EE is the entry ticket for every AI feature that matters, and every option discussed below is priced on top of that base. Read our guide to the 23ai Free versus Enterprise Edition feature line before you commit an edition to any AI project.
Here is the reversal. The AI feature is free, but the options that surround real 23ai deployments are not, and several of them switch on through ordinary use rather than through an explicit purchase decision. The table below lists the 2026 US list prices for the options most likely to appear in a 23ai audit finding. Treat these as pre-discount reference points, not as what you should pay.
| Option or Pack | US list per processor | US list per NUP | How it switches on in 23ai |
|---|---|---|---|
| Multitenant | $17,500 | $350 | 4th pluggable database in a container database |
| Partitioning | $11,500 | $230 | Partitioned tables, including IVF vector indexes at scale |
| Real Application Clusters (RAC) | $23,000 | $460 | Clustered deployment for vector or OLTP workloads |
| Spatial and Graph | $17,500 | $350 | Specific PL/SQL package execution, often via application code |
| Diagnostics Pack | $7,500 | $150 | One AWR query or Enterprise Manager click |
| Tuning Pack | $5,000 | $100 | SQL Tuning Advisor use, often via Enterprise Manager |
| Advanced Security | $15,000 | $300 | TDE, redaction, or related security use |
| Database Vault | $11,500 | $230 | Realm or command rule configuration |
The Diagnostics and Tuning packs are the packs found most often in Oracle audits, because they light up through Enterprise Manager and through routine performance queries. A single Automatic Workload Repository (AWR) query records usage you never intended to buy. These are not 23ai-specific problems, but 23ai deployments run through Enterprise Manager like everything else, and the AI marketing push tends to draw in teams who have never managed Oracle license compliance before. The Cloud Management Pack self-enablement problem is a related trap on the same Enterprise Manager surface.
Oracle's feature-usage view keeps a permanent record of every option ever touched, including a single test query from two years ago. Auditors read it first.
Multitenant deserves its own section because 23ai adoption tends to multiply pluggable databases. Enterprise Edition includes up to three pluggable databases (PDBs) per container database (CDB) at no extra charge. The fourth PDB triggers Multitenant licensing (17,500 US dollars per processor, 350 US dollars per NUP) for the entire CDB, and for every container that hosts more than three PDBs. This is one of the most common and most expensive traps in the entire Oracle Database portfolio.
Why does 23ai make this worse? Teams building AI and vector workloads tend to spin up separate PDBs for each model, each embedding pipeline, and each RAG application, because PDBs are lightweight and easy to provision. It feels free because the first three are free. The moment a busy consolidation platform crosses three PDBs in a single CDB, the whole container is licensable for Multitenant across every processor it runs on. On a large Exadata or engineered-systems footprint, that single architectural decision can add six or seven figures of exposure.
The defensible position is deliberate PDB governance: cap production CDBs at three PDBs unless you have Multitenant licensed with headroom, and treat any fourth-PDB request as a purchasing decision that goes through licensing review, not through a DBA ticket. If you are consolidating aggressively, license Multitenant on purpose and document it, rather than tripping into it. Our CIO playbook on option pack rationalization covers the governance controls in depth.
This is the 23ai-specific exposure most buyers miss. AI Vector Search is free, but the way you scale it can pull in two of the most expensive options Oracle sells. The HNSW (Hierarchical Navigable Small World) index, which delivers the fastest approximate nearest-neighbor search, is an in-memory, instance-local structure. It is not supported on Real Application Clusters. On a clustered database you must use the Inverted File Flat (IVF) index instead, which partitions vectors and works across instances.
Notice the word 'partitions'. Oracle's own documentation states that features like RAC and Partitioning work seamlessly for vector data. Read that as a buyer, not as a marketer: if you run vector search on a clustered production database for high availability or scale, you will use RAC (23,000 US dollars per processor list) and quite possibly Partitioning (11,500 US dollars per processor list) to make IVF vector indexes perform. The 'free' vector feature has just become a 34,500-dollar-per-processor list conversation before discount, and that is per processor across the whole cluster.
The HNSW index carries its own quieter cost driver. HNSW indexes live in memory and require you to configure the VECTOR_MEMORY_SIZE parameter to carve out a dedicated vector pool in the System Global Area (SGA). That is not itself a licensable option, but it raises memory pressure that pushes some buyers toward Database In-Memory or larger engineered-systems configurations, both of which have license consequences. The point is not that any single step is a trap. The point is that the free feature sits at the center of a web of priced features, and production-grade vector search touches several strands of that web at once.
Your action: architect vector workloads on single-instance Enterprise Edition wherever high availability requirements allow, using HNSW in-memory indexes. Move to RAC and IVF only when a documented availability or scale requirement forces it, and license those options deliberately with negotiated discount rather than discovering them in an audit. See our deep dive on where the vector search license line actually sits.
To generate vector embeddings inside the database rather than calling an external service, 23ai lets you load transformer models as first-class database objects. Oracle Machine Learning enables loading text embedding models in ONNX format directly into the database using the PL/SQL packages DBMS_VECTOR and DBMS_DATA_MINING (the latter from OML4SQL). This is a genuinely useful capability: your embedding pipeline runs where your data lives, with no data leaving the database.
The licensing question is sharper than Oracle's marketing suggests. DBMS_VECTOR is presented as part of the free AI Vector Search feature set. DBMS_DATA_MINING originates in Oracle Machine Learning for SQL, which historically sat under the Oracle Advanced Analytics option, later folded into Enterprise Edition. In current 23ai positioning, the in-database machine learning packages are described as included in Enterprise Edition, but the boundary between 'included ML for vector embeddings' and 'the broader machine learning and data mining functionality' is exactly the kind of gray zone Oracle audits probe.
Our reading, based on market experience rather than a single definitive Oracle statement, is that embedding generation through DBMS_VECTOR sits inside the free AI Vector Search feature, while heavier use of DBMS_DATA_MINING for predictive modeling and classification is a different conversation that you should confirm in writing. If your teams start using OML4SQL for anything beyond loading embedding models, get a written scope statement from Oracle. We treat this boundary in detail in our analysis of the 23ai in-database machine learning and Advanced Analytics license question.
True Cache is included in Enterprise Edition, but it comes with a licensing rule that buyers routinely miss. The 23ai licensing manual has been clarified to state that all True Cache instances must match the edition of the primary database. If your primary runs Enterprise Edition, your True Cache instance must be Enterprise Edition too, and it must be licensed on the hardware it runs on. True Cache does not create a second free database license; it creates a second EE deployment you must count.
There is a subtler point that sits unresolved in the documentation. True Cache uses Active Data Guard technology under the covers, being a read-only replica of the primary. Active Data Guard is a separately licensed option (part of the Advanced options family). One licensing analysis observed that True Cache uses Active Data Guard technology but noted that the licensing manual expresses no explicit obligation to license Active Data Guard for True Cache. This is a live ambiguity. Oracle could argue either way in an audit.
True Cache is bundled, but it is a second Enterprise Edition deployment on real hardware, and its Active Data Guard lineage is an unresolved audit question.
Buyer action: license the True Cache host for Enterprise Edition, and get a written confirmation from Oracle that True Cache does not require a separate Active Data Guard license. Do not accept a verbal assurance. Put it in the ordering document or in a signed clarification, because the gap between 'the manual is silent' and 'you owe us Active Data Guard' is where audit claims are born. Our dedicated piece asks directly whether True Cache is a free feature or a second database.
SQL Firewall is a 23ai security feature that inspects SQL at the kernel level to block unauthorized statements. Here the sources genuinely conflict, and you need to know that before you deploy it. One authoritative licensing analysis states that SQL Firewall has been clarified as part of Database Vault or Audit Vault and Database Firewall, both of which are separately licensed paid options. A separate vendor comparison claims SQL Firewall is incorporated directly into the database kernel and included in both editions.
Do not resolve this conflict by choosing the cheaper interpretation. Resolve it by treating SQL Firewall as licensable until Oracle confirms otherwise in writing for your specific version and edition. If a security team enables SQL Firewall on the assumption it is free, and Oracle's audit position is that it belongs to Database Vault (roughly 11,500 US dollars per processor list) or the Advanced Security family (roughly 15,000 US dollars per processor list), you inherit a per-processor liability across every server that touched the feature. The feature-usage view will have recorded it either way.
Advanced Security, Database Vault, and Label Security all carry separate per-processor costs, published in the range of 11,500 to 15,000 pounds sterling per processor. Any 23ai security hardening project should treat every security feature as potentially licensable and confirm each one against your specific ordering document, not against a marketing comparison chart.
Every point above converges on one mechanism. Oracle's automatic feature-tracking view (DBA_FEATURE_USAGE_STATISTICS and related views) captures every option ever used on every database, including a single test query that ran two years ago. The view retains usage history for the life of the database. Oracle LMS auditors examine this view as standard practice, before they look at anything else.
This is why 23ai raises audit exposure even though its flagship feature is free. New AI projects bring new teams, new PDBs, new indexes, new PL/SQL packages, and new Enterprise Manager activity into your Oracle estate. Each of those actions can flip a feature flag in the usage view. A developer running SQL Tuning Advisor to speed up a slow vector query records Tuning Pack usage. A team crossing the fourth PDB records Multitenant usage. An architect enabling Partitioning to scale IVF indexes records Partitioning usage. None of it feels like a purchase. All of it is permanent, and all of it is discoverable.
The defensive discipline is straightforward and non-negotiable. Before you upgrade to 23ai, run the feature-usage view across your estate and capture a baseline. After you deploy AI features, run it again and diff the results. Anything new that lit up is either something you licensed on purpose or something you need to remediate immediately, ideally by disabling the feature and documenting the date. Our overview of the Oracle license audit process explains how LMS builds a claim from exactly this data, and how buyers cut the number.
Upgrading to 23ai is not a like-for-like version bump from a licensing perspective. It is a moment when new features become available and new usage becomes possible. Run a structured review before the migration, not after the audit letter arrives.
The single highest-value action on that list is the written confirmation of the three ambiguous items. Ambiguity always resolves in Oracle's favor during an audit, and the cost of getting three sentences into your ordering document before deployment is zero. The cost of losing the argument afterward is measured per processor across your entire vector-search footprint. Follow our step-by-step 23ai license review to run before you migrate.
One escape route from the options maze is to run 23ai AI features on Autonomous Database rather than self-managed Enterprise Edition. On Autonomous, many options that are separately licensed on-premises (Partitioning, RAC, Advanced Security, Diagnostics, Tuning) are bundled into the consumption price. That removes the feature-usage-view audit risk for those specific options, because you are paying by consumption rather than by option flag.
That is not a free lunch. Autonomous shifts you to a metered cloud cost that can exceed a well-managed on-premises EE deployment for steady, predictable workloads, and it locks you into Oracle Cloud Infrastructure. The decision is genuinely workload-dependent: bursty, experimental AI and RAG workloads often favor Autonomous, while steady high-throughput production vector search on owned hardware often favors self-managed EE with deliberately licensed options. Model both before you commit, and read our comparison of running 23ai AI features on Autonomous versus self-managed.
Consider a mid-sized buyer who upgrades a two-node RAC cluster (assume 16 processor licenses total under Oracle's counting) to 23ai to run production vector search with high availability. They already own Enterprise Edition. Vector search itself costs nothing extra. But to run vectors across the cluster they use IVF indexes, which need Partitioning, and the cluster itself needs RAC. They spin up a fourth PDB for the embedding pipeline, triggering Multitenant on the whole container. A developer runs SQL Tuning Advisor, lighting up the Tuning Pack.
| Trigger | Option | List per processor (USD) | x16 processors (list) |
|---|---|---|---|
| Clustered deployment | RAC | $23,000 | $368,000 |
| IVF vector indexes at scale | Partitioning | $11,500 | $184,000 |
| 4th PDB in the CDB | Multitenant | $17,500 | $280,000 |
| SQL Tuning Advisor use | Tuning Pack | $5,000 | $80,000 |
| Total incremental list exposure | $912,000 |
That is 912,000 US dollars of list-price exposure created by a project whose flagship feature was advertised as free. The numbers are illustrative and pre-discount, and your negotiated rates will be lower, but the structure is exactly what we see in real 23ai deployments. The lesson is not that 23ai is expensive. The lesson is that the free AI feature and the priced infrastructure around it must be planned as one budget, not as a free experiment that quietly grows teeth. For the broader cost picture across metrics, see our overview of Oracle licensing cost in 2026.
Three moves, in order. First, baseline your feature-usage view before any 23ai migration and rerun it after each deployment phase. This single control catches the majority of accidental option activation and gives you a documented remediation timeline if something lights up. Second, get written confirmation on the three ambiguous items (True Cache versus Active Data Guard, SQL Firewall edition inclusion, DBMS_DATA_MINING scope) inside your ordering document before you deploy, because ambiguity always resolves in Oracle's favor during an audit.
Third, architect deliberately. Keep AI workloads on single-instance Enterprise Edition with HNSW in-memory indexes wherever availability requirements permit, cap PDBs at three per CDB, and treat any move to RAC, Partitioning, or Multitenant as a purchasing decision with negotiated discount rather than a technical afterthought. The free feature is real and worth using. The discipline around it is what keeps it free. Run your footprint review well before renewal so any 23ai exposure is remediated on your timeline, not Oracle's.
Yes. AI Vector Search, including the VECTOR data type, HNSW and IVF indexes, and nearest-neighbor search, is included in Enterprise Edition at no additional license cost. It also ships in Oracle Database Free. Standard Edition 2 cannot run it, so you need Enterprise Edition as the base.
Scaling vector search across a cluster pulls in Real Application Clusters (about 23,000 USD per processor list) and Partitioning (about 11,500 USD list). A fourth pluggable database triggers Multitenant across the whole container database. Diagnostics and Tuning packs light up through routine Enterprise Manager use. None of these feel like purchases, but all are recorded permanently in the feature-usage view.
True Cache is included in Enterprise Edition, but all True Cache instances must match the edition of the primary database and be licensed on their own hardware. It uses Active Data Guard technology under the covers, and the licensing manual is silent on whether Active Data Guard must be licensed separately. Get written confirmation from Oracle before you deploy it.
Oracle's feature-usage view captures every option ever used and retains that history for the life of the database, and LMS auditors examine it first. New AI projects bring new PDBs, indexes, PL/SQL packages, and Enterprise Manager activity that can each flip an option flag. Baseline the view before migration and rerun it after each phase to catch accidental activation.
No. SE2 cannot run AI Vector Search, RAC, Partitioning, Advanced Security, or Multitenant beyond three PDBs. Any native vector workload requires Enterprise Edition. Isolate the AI workload on EE rather than upgrading an entire SE2 estate you do not otherwise need to upgrade.
Embedding generation through DBMS_VECTOR is positioned inside the free AI Vector Search feature. Heavier use of DBMS_DATA_MINING (from Oracle Machine Learning for SQL) for predictive modeling sits in a grayer zone historically tied to the Advanced Analytics option. Confirm the scope in writing if your teams use OML4SQL for anything beyond loading embedding models.
Oracle Database 23ai bundles options you may never deploy. The buyer side guide to edition right sizing, option pruning, and AI Vector Search licensing.
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.
