AI Vector Search is free in Enterprise Edition, but the platform you run it on decides whether you pay for options separately or get them bundled. This guide quantifies both paths and tells you where the crossover sits.
AI Vector Search is free in Enterprise Edition, but the platform you run it on decides whether you pay for options separately or get them bundled. This guide quantifies both paths and tells you where the crossover sits.
Oracle's marketing line is technically accurate and commercially misleading in equal measure. AI Vector Search (the native VECTOR data type, HNSW and IVF indexing, and approximate and exact nearest neighbor search) ships inside Oracle Database Enterprise Edition at no additional license cost, and it is even present in the free tier. That is the fact buyers latch onto, and it is the fact Oracle account teams repeat. What it hides is that the vector feature is a passenger, not the vehicle. Whether it costs you nothing or a great deal depends entirely on the platform underneath it and the options that platform forces you to license.
The decision you are actually making is not 'do I pay for AI Vector Search.' It is 'do I consume it on Autonomous Database, where every Enterprise Edition option is bundled into the hourly rate, or do I consume it on self-managed Enterprise Edition, where each option that vector workloads touch (Partitioning, RAC, and the diagnostic packs) is separately licensed and separately audited.' That framing is different from the standard Autonomous cost analysis. It is the framing that saves money, and we cover the underlying feature-inclusion question in more depth in our companion piece on whether 23ai vector search is included in Enterprise Edition or an extra.
One naming note before the numbers, because it trips up 2026 procurement teams: Oracle renamed the release. Oracle AI Database 26ai replaces 23ai, and the transition is a Release Update (October 2025), not a database upgrade or application re-certification. On-premises Enterprise Edition for Linux x86-64 arrived in the January 2026 quarterly Release Update (version 23.26.1). Everything in this cost comparison applies to both labels; the licensing mechanics are unchanged.
On Autonomous Database with License Included, Oracle bundles every option the service supports into the consumption rate. That list is not trivial: Transparent Data Encryption, Multitenant, RAC, Partitioning, Advanced Compression, Advanced Security, Label Security, Database Vault, and Automatic Data Guard are all included, none separately licensed. For a vector workload, the two that matter most are Partitioning (because IVF vector indexes support local partitioning) and RAC (for scale-out). On Autonomous you never see a line item for either. That is the structural advantage of the platform and the single strongest argument for it when your AI features touch partitioned tables.
The pricing metric is the ECPU, consumption-based rather than perpetual. At list, an Autonomous Database ECPU runs approximately $0.336 per hour with license included, per current 2026 rate cards. The published band runs wider: from $0.4032 per OCPU hour for Autonomous Transaction Processing Serverless up to $2.86 per OCPU hour for Dedicated Exadata. ECPU instances require a minimum of 2 ECPUs, and Autonomous Serverless storage is priced separately. The smallest configuration is 2 ECPUs, and there is a fixed-shape Autonomous AI Database for Developers (4 ECPU with 20 GB storage) priced hourly per instance for proof-of-concept work.
| Autonomous configuration | Metric | List rate | Notes |
|---|---|---|---|
| ATP Serverless | OCPU/hour | $0.4032 | Entry serverless tier |
| Standard ECPU | ECPU/hour | ~$0.336 | Minimum 2 ECPUs |
| Dedicated Exadata | OCPU/hour | $2.86 | High-end, storage offload |
| Developer fixed shape | 4 ECPU / 20 GB | Hourly per instance | POC and dev only |
Two levers materially change these numbers before you sign. First, annual commit deals typically clear at 25 to 55 percent discount to published rates; treat list as a ceiling, never a starting point. Second, Elastic Pools let you consolidate autonomous databases into a shared pool for up to 87 percent compute cost savings when your instance fleet is bursty rather than steady-state. If you are running many small vector databases, Elastic Pools is the difference between an economical platform and an expensive one. We walk the full metric and math in the Autonomous Database licensing guide.
On Autonomous you never license Partitioning or RAC separately. On self-managed, a partitioned vector index makes both a compliance exposure.
Self-managed Enterprise Edition on your own infrastructure or on OCI compute gives you AI Vector Search at no incremental feature cost, exactly as advertised. The cost lives elsewhere. Enterprise Edition itself is priced on the Processor metric, referenced in 2025 European pricing at approximately GBP 47,500 per processor, and running roughly 2.7 times Standard Edition 2 per Processor and per Named User Plus. Note that AI Vector Search is not available in Standard Edition 2 at all, so the cheaper edition is off the table the moment you commit to native vector search. That fact alone forces many buyers up to EE who assumed SE2 would suffice.
The traps sit in the options that vector workloads pull in. IVF vector indexes support local partitioning, so any self-managed customer indexing partitioned vector tables must own the separately licensed Partitioning option. RAC, if you want scale-out for millions of embeddings, is another separately licensed option in the self-managed world. Neither appears on your invoice until an audit finds the feature usage, at which point back-support and list-price penalties apply. This is the recurring pattern in Oracle self-managed deployments: the feature is free, the option that makes the feature useful is not, and the gap surfaces during a review. Our 23ai licensing guide catalogs which options moved into base and which still trigger a bill.
There is also a production-readiness dimension that changes the self-managed calculus specifically. In 23ai, the HNSW index did not fully support DML on tables carrying an HNSW index on the vector column. 26ai fixes this, meaning vector search queries now see transactional consistency on indexed tables. If you standardized on 23ai self-managed early, you inherited a production limitation that Oracle has since resolved by Release Update. Confirm you are on 23.26.1 or later before benchmarking self-managed against Autonomous, or you will compare the wrong product.
The two paths are not mutually exclusive. If you already hold perpetual Enterprise Edition licenses, Bring Your Own License lets you carry them onto Autonomous and pay only for infrastructure, not the software rate. This is the single largest cost reduction available on the platform. One source quantifies BYOL as reducing Autonomous compute costs by 76 percent versus License Included; another describes the rate cut as roughly 40 percent. The spread reflects different configurations and discount baselines, so model your own numbers rather than trusting either figure blindly.
The conversion ratio is where you must be careful, because published sources conflict. One Oracle blog states that one Processor license (or 25 Named User Plus licenses) covers eight ECPUs or two OCPUs of Autonomous Database. A separate 2026 account states one on-premises Processor license now equals two ECPUs of BYOL credit. These are materially different ratios. Do not accept a verbal figure from your account manager. Pull the current OCI Service Description in force at your contract date and use that exact ratio in writing. Getting this wrong by a factor of four is a real compliance and budget risk.
| BYOL element | What it delivers | Buyer action |
|---|---|---|
| Rate reduction | 40% to 76% off License Included | Model your own workload, do not trust the headline |
| Conversion ratio | Sources conflict (2:1 vs 8:1 ECPU) | Confirm against current OCI Service Description |
| Support prerequisite | Licenses must be in active support | Verify support renewal before counting BYOL credit |
| Bonus option rights | TDE, Diagnostics, Tuning, Data Masking, RAT included | Claim these; do not re-buy separately |
| Peak measurement | Must license peak ECPU consumed | Cap auto-scaling to control license exposure |
Two prerequisites gate the BYOL discount. First, the perpetual licenses must be in active support. Lapsed support kills the BYOL credit, so a customer who dropped support to save annual fees cannot later resurrect the entitlement for Autonomous without re-instatement penalties. Second, and this is the audit trap, you must hold sufficient licenses for the peak ECPU consumed, not the average. Auto-scaling that spikes to 32 ECPUs at month-end obligates you to license 32 ECPUs of entitlement even if your steady state is 8. Cap auto-scaling deliberately, because on Autonomous the elasticity that Oracle sells as a benefit becomes your compliance liability under BYOL.
BYOL carries two upsides worth naming. When you bring an Enterprise Edition entitlement, Oracle additionally grants rights to TDE, Diagnostics Pack, Tuning Pack, Data Masking and Subsetting Pack, and Real Application Testing without bringing separate entitlements for those options. That is genuine value; do not let anyone sell you those packs a second time. And since 2024, Autonomous Serverless lets you mix BYOL ECPUs and License Included ECPUs in the same instance or elastic pool by assigning the billing method per usage level, so you can cover your baseline with owned licenses and burst on License Included.
Confirm the BYOL ECPU ratio in writing from the current OCI Service Description. Sources disagree by a factor of four, and that gap is your budget.
Both paths share the same engine, so both share the same scale ceiling against standalone vector databases. AI Vector Search is genuinely competitive with dedicated vector platforms at moderate scale (under 10 million vectors). Above that, standalone platforms retain an advantage on raw query latency. If your embedding corpus is heading past 10 million vectors, the Autonomous-versus-self-managed question is partly moot; the real comparison is Oracle versus a purpose-built vector store, and that changes the analysis entirely.
One caveat favors Autonomous and Exadata at the high end. Oracle Exadata for AI offloads vector search to intelligent storage for significant speedups, and the newer Exascale architecture extends this to smaller, lower-cost deployments. This storage offload is an Autonomous and Exadata advantage that self-managed commodity hardware cannot replicate. If your workload is both large and latency-sensitive, the platform premium for Autonomous on Dedicated Exadata may pay for itself, but note that Autonomous on Dedicated Exadata with BYOL requires you to hold Enterprise Edition plus the RAC option for the relevant components. That is a real entitlement requirement, not an optional one.
For a workload under 10 million vectors with steady, predictable compute and no existing license inventory, Autonomous License Included is the cleaner choice because it removes the Partitioning and RAC option exposure entirely and requires no capacity planning. For an organization with a stock of Enterprise Edition licenses in active support, BYOL onto Autonomous almost always wins on cost, provided you cap auto-scaling and confirm the conversion ratio. For a fully self-managed shop with existing DBA capacity and on-premises data residency requirements, self-managed EE is defensible only if you deliberately license Partitioning and RAC where vector indexes and scale-out demand them, and only after you have run the option-usage review.
Whichever path you choose, run a license review before you migrate any AI workload, because the vector feature quietly activates options you may not be tracking. Our checklist for that exercise is in the 23ai upgrade license review, and the broader trigger analysis lives in the 23ai vector and AI feature licensing pillar. For the underlying Enterprise Edition cost math on the self-managed side, see our 2026 Oracle Database license cost breakdown.
Yes, the feature itself ships in Enterprise Edition and the free tier at no additional license cost. The cost lives in the platform and the options vector workloads touch, chiefly Partitioning and RAC on self-managed deployments, which are bundled on Autonomous but separately licensed on self-managed Enterprise Edition.
Published sources put the reduction between 40 percent and 76 percent versus License Included pricing, with the range depending on configuration and discount baseline. You must model your own workload, and the licenses you bring must be in active support to qualify.
Sources conflict. One Oracle blog states one Processor license covers eight ECPUs or two OCPUs; a 2026 account states one Processor equals two ECPUs. This is a factor-of-four discrepancy, so confirm the exact ratio against the current OCI Service Description before you sign anything.
No. AI Vector Search is not available in Standard Edition 2, alongside RAC, Partitioning, Advanced Compression, and other options. Native vector search forces you onto Enterprise Edition, which runs roughly 2.7 times SE2 per Processor.
Yes. Under BYOL you must hold licenses for the peak ECPU consumed, not the average. Uncapped auto-scaling that spikes at month-end obligates you to license that peak. Cap auto-scaling to your licensed ECPU count to control exposure.
AI Vector Search is competitive under 10 million vectors. Above that, standalone vector databases retain a raw query latency advantage, though Exadata storage offload narrows the gap for Autonomous and Exadata deployments. Benchmark both before committing.
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.
