Oracle Select AI — How It's Licensed and What It Costs to Run

Oracle Select AI licensing is the commercial wrapper around the SQL-native generative AI capability that Oracle ships with Autonomous Database. The SELECT AI clause and the DBMS_CLOUD_AI PL/SQL package allow a developer to issue a natural-language question against the database and receive a generated SQL statement, a query result, a narrated response, or a chat-style continuation — all inside the database session, without leaving the SQL layer. The headline licensing position is that Select AI is bundled with Autonomous Database. The economic question is what it costs to run at production scale.

This piece works the Oracle Select AI licensing position the way an Oracle insider building the consumption forecast for a customer-service business-user rollout would work it: the licensing model first, the two-layer cost mechanics second, the provider routing decision third, the supporting feature integration fourth, and the buyer-side commercial framework last. For the broader Autonomous Database context see the Oracle Cloud licensing master guide; for the related vector search capability see our Oracle 23ai AI Vector Search licensing analysis.

The Oracle Select AI licensing model

Bundled with Autonomous Database

Oracle Select AI is shipped as a feature of Oracle Autonomous Database through the DBMS_CLOUD_AI PL/SQL package and the SELECT AI SQL clause syntax. The functionality is available on Autonomous Database 23ai (the current AI-focused release) and on supported maintenance levels of Autonomous Database 19c. The customer does not pay a separate Select AI option licence — the capability is included in the consumption-based Autonomous Database pricing model at the published ECPU-hour rate.

Self-managed Oracle Database — no native Select AI

Self-managed Oracle Database 23ai deployments do not have the DBMS_CLOUD_AI package built into the standard distribution in the same way Autonomous does. Customers running self-managed Enterprise Edition deployments who want equivalent functionality have three options. Option one is implementing the pattern manually using UTL_HTTP to call the chosen generative AI provider API from PL/SQL — feasible but requires the application developer to build the NL2SQL translation, the safety checks, and the prompt engineering scaffolding. Option two is using Oracle Application Express AI components, which provide a managed layer over the same pattern. Option three is using a middleware integration layer (LangChain, LlamaIndex, or a similar framework) outside the database. The licensing implications differ across the three options — Autonomous bundles it, self-managed does not.

The provider routing decision

Select AI is provider-agnostic — the AI profile object specifies which generative AI provider the call routes to. The supported providers include Oracle Generative AI Service (the default for OCI-native deployments), OpenAI (GPT-4, GPT-4 Turbo, GPT-4o), Anthropic (Claude family), Cohere (Command R, Command R+, Embed), Azure OpenAI Service, Google Vertex AI, and Hugging Face hosted models. The per-token rate billing happens at the provider — the customer's Select AI deployment can route to OCI Generative AI Service against Universal Credits, or to an external provider against a separate commercial commitment, or to multiple providers across different AI profiles in the same Autonomous instance.

The two-layer cost mechanics

Layer 1 — Autonomous Database ECPU consumption

The first cost layer is the Autonomous Database ECPU-hour consumption that hosts the SQL session and processes the Select AI call. At the published 2026 rate of $0.336 per ECPU-hour, a single-ECPU Autonomous Database instance running 24x7 lands at approximately $245 per month per ECPU. A production deployment supporting business-user query traffic typically requires 2-8 ECPUs depending on the concurrency profile — landing somewhere between $490 and $1,960 per month for the database hosting layer alone. The Select AI call itself is a relatively lightweight database operation (a PL/SQL call to DBMS_CLOUD_AI, an external HTTP request to the provider, the response parsing) — the ECPU consumption per query is typically negligible compared with the provider call.

Layer 2 — Generative AI provider per-token consumption

The second cost layer is the per-token consumption at the chosen generative AI provider. The Select AI call sends a prompt to the provider containing the natural-language question, the schema metadata for the relevant database objects, the example query patterns, and any retrieval-augmented context (vector search results, narrative templates). The prompt size is typically 1,500-4,000 input tokens depending on the schema complexity and the prompt engineering. The response is typically 200-800 output tokens depending on whether the call is using runsql (just the SQL), narrate (the SQL plus a narrated explanation), or chat (a multi-turn conversational continuation).

At 100,000 Select AI queries per month against Cohere Command R, the provider cost lands at $80-250 per month — modest. At 1 million queries per month against GPT-4o, the provider cost lands at $2,000-14,000 per month — material. The forecast has to model the projected query volume against the chosen provider at the actual prompt size, not the headline per-query rate.

The provider routing decision

Oracle Generative AI Service — the OCI-native default

Routing Select AI calls to Oracle Generative AI Service keeps the consumption inside the Universal Credits commitment and benefits from the commitment-based discount tiering. The trade-off is the model catalogue — OCI Generative AI Service offers Cohere and Meta Llama variants, not GPT-4 or Claude. For NL2SQL workloads against well-structured schemas the Cohere Command R+ model is typically sufficient; for complex multi-table queries or conversational continuations the higher-end models may produce better results.

OpenAI and Azure OpenAI — the GPT-4 path

Routing Select AI calls to OpenAI or Azure OpenAI gives access to the GPT-4 family for the NL2SQL translation. The consumption bills against the OpenAI or Azure commercial commitment rather than the Oracle Universal Credits position. For customers with a material Azure OpenAI commitment under an existing Microsoft Enterprise Agreement, the Azure OpenAI route frequently lands at a lower effective per-token rate than the published OCI Generative AI rate.

Anthropic — the Claude path

Routing Select AI calls to Anthropic gives access to the Claude family. The Claude models typically produce strong NL2SQL results on complex schemas and well-documented business questions. The consumption bills against the Anthropic commercial commitment.

The multi-provider AI profile pattern

A material buyer-side architectural pattern is to configure multiple AI profiles inside the same Autonomous instance — one for fast lightweight queries against Cohere Command R, one for high-quality complex queries against GPT-4 or Claude, one for embedding generation against Cohere Embed or OpenAI Ada. The Select AI call selects the profile by name. This pattern allows the customer to right-size the per-query cost against the per-query value rather than routing every call to the highest-cost provider.

The supporting feature integration

Select AI + AI Vector Search — the RAG pattern

Select AI integrates with AI Vector Search in 23ai to support retrieval-augmented generation patterns. The typical RAG pattern uses a Select AI call with an AI profile pointing at an embedding model to convert the natural-language question into a vector, queries the vector index for the most semantically relevant context, then issues a second Select AI call with the retrieved context augmenting the prompt. The licence implication is that both Select AI and AI Vector Search consume from the same Autonomous Database ECPU bill — the cost compounds. For deployment patterns combining the two see our RAG architecture on Oracle Database licensing analysis.

Select AI + APEX

Oracle Application Express integrates with Select AI through the APEX Assistant components — the application developer can deploy a natural-language interface against an APEX application backed by Autonomous Database. The licensing pattern is the same as standalone Select AI: APEX runs free with Autonomous, the Select AI calls bill against the two-layer cost model.

Select AI + Oracle Analytics Cloud

Oracle Analytics Cloud has its own natural-language query capabilities (Oracle Analytics Day by Day, the OAC assistant). Customers running Select AI on the underlying Autonomous Database and OAC on top of it should explicitly model which layer the NL query routes through — Select AI in the database produces different commercial implications than the OAC NL feature, and routing every business-user question through both creates duplicate provider consumption.

"Oracle Select AI is the cleanest SQL-native generative AI capability shipped by any major database vendor. The bundle with Autonomous Database is real. The buyer-side defence is to model the two-layer cost — Autonomous ECPU plus the underlying provider — and to route queries to the right-priced provider rather than letting the application team default everything to the highest-cost model."

An anonymised case study — global insurer, Select AI business-user rollout

A global insurer with a 16-ECPU Autonomous Database 23ai footprint planned a Select AI rollout to 1,200 business-user analysts in early 2026. The original deployment proposed by the Oracle account team configured Select AI with a single AI profile routing to Oracle Generative AI Service Cohere Command R+. The projected consumption — based on 50 queries per user per day at the Cohere Command R+ list rate — landed at $84k per month of provider consumption against the Universal Credits commitment.

The buyer-side architecture review modelled three options. Option A was the proposed single-profile deployment at $84k per month. Option B was a multi-profile architecture routing 80% of queries (simple lookups, standard reports) to Cohere Command R, 15% (complex analytical queries) to Cohere Command R+, and 5% (multi-table exploratory queries) to Anthropic Claude Sonnet via a separate commercial commitment — projected at $34k per month total. Option C was a hybrid pattern with Azure OpenAI GPT-4o as the primary provider for 90% of queries (the existing Microsoft Enterprise Agreement covered the consumption at preferential rates) and OCI Generative AI Service as the fallback for the remaining 10% — projected at $22k per month total.

The buyer-side recommendation was Option B with three commercial provisions. First, the OCI Generative AI Service consumption was renegotiated to absorb against the existing Universal Credits commitment at the 35% commitment discount tier, bringing the Cohere Command R+ effective rate down to the level where the per-query economics worked for the complex analytical workload. Second, the Anthropic Claude Sonnet consumption was capped with a written commercial provision against the Anthropic commitment — protection against runaway expensive queries. Third, the AI profile configuration was governed at the database level with explicit role-based controls — the application team could not route every query to the highest-cost model. Net annualised consumption: $408k. Saving against the original Oracle proposal: $600k per year. For the broader Oracle commercial framework see our Oracle contract negotiation service.

The five buyer-side moves on Oracle Select AI

Move 1 — Model the two-layer cost explicitly. The Autonomous ECPU bill is the predictable component. The provider per-token bill is the variable component and frequently the dominant cost at scale. Forecast both layers before the deployment business case.

Move 2 — Right-size the provider routing. Configure multiple AI profiles in the same Autonomous instance — fast cheap models for simple queries, premium models for complex queries. Avoid routing every business-user question through the highest-cost model.

Move 3 — Negotiate the Universal Credits absorption. If Select AI routes to Oracle Generative AI Service, the per-token cost draws against the Universal Credits commitment with commitment-based discount tiering. Treat the consumption as part of the broader OCI commercial conversation.

Move 4 — Govern the AI profile configuration at the database level. Role-based controls on which application teams can use which AI profiles prevent the application layer from defaulting every call to the most expensive provider. The governance is the cost control.

Move 5 — Cap the provider commitment. Whichever provider Select AI routes to, the commercial commitment with the provider should include a written ceiling on monthly consumption with notification triggers. Oracle's playbook on consumption-based services is to let the bill grow without notice — the buyer-side defence is the contractual cap with monitoring.

Frequently asked questions

How is Oracle Select AI licensed?

Oracle Select AI is licensed as a feature of Oracle Autonomous Database — the SELECT AI SQL function and the DBMS_CLOUD_AI package are bundled with the Autonomous Database service at no separate uplift. The cost model is two-layered. Layer one is the Autonomous Database ECPU-hour consumption that hosts the database and processes the SQL statement. Layer two is the underlying generative AI service consumption — the Select AI call invokes a foundation model on Oracle Generative AI Service, OpenAI, Cohere, Anthropic, Google Vertex AI, or an Azure OpenAI endpoint, with the per-token cost billing against the chosen provider. The customer pays for both layers.

Can I use Oracle Select AI on Oracle Database 23ai self-managed?

The DBMS_CLOUD_AI package that powers Select AI is currently shipped as a feature of Oracle Autonomous Database (Autonomous Database 23ai, Autonomous Database 19c with the appropriate maintenance level). Self-managed Oracle Database 23ai deployments do not have the native Select AI capability built into the same DBMS_CLOUD_AI package — customers wanting equivalent functionality on self-managed deployments typically implement the pattern manually using UTL_HTTP to call the generative AI provider API, or use Oracle Application Express AI components, or use a middleware integration layer.

What does Oracle Select AI cost per query?

The cost per Select AI query is the sum of three components. Component one is the Autonomous Database ECPU consumption for the SQL execution, typically negligible per query at the published $0.336 per ECPU-hour rate. Component two is the foundation model per-token cost for the natural language understanding (NL2SQL translation) — typically 200-500 input tokens and 100-200 output tokens per query at the chosen model rate. Component three is the foundation model per-token cost for the response narration if the call uses the narrate or chat action — additional 100-500 tokens depending on the response complexity. At indicative 2026 rates a typical Select AI query lands at $0.001 to $0.005 per query against Oracle Generative AI Service Cohere Command R, or $0.002 to $0.012 per query against OpenAI GPT-4.

Does Oracle Select AI trigger an audit exposure?

Oracle Select AI runs entirely on Autonomous Database with bundled DBMS_CLOUD_AI functionality, so the feature itself does not generate a separate Oracle Database option audit exposure in the way an Enterprise Edition deployment using Partitioning without an option licence would. The exposure profile is different — the consumption forecast across both Autonomous ECPU and the underlying generative AI provider can grow rapidly when Select AI is deployed at scale across business-user workloads, and the deployment frequently activates supporting components (AI profiles pointing at multiple providers, RAG pipelines combining Select AI with AI Vector Search) that compound the licence exposure on the database side.

The broader AI estate Select AI sits inside

Select AI is a narrow SQL-native surface, but the audit exposure widens whenever it is paired with the retrieval and orchestration patterns most customers build around it. Customers running Select AI with 23ai Vector Search as the retrieval backend should defend the option-attach exposure forensic in our RAG on Oracle Database licensing analysis, and audit the third-party framework wiring against our LangChain Oracle licensing risk analysis. Customers absorbing Select AI into a broader AI commitment with Oracle should benchmark the bundled-platform commercial pressure forensic in our Oracle AI Data Platform licensing analysis, and right-size the conversational reach using our Oracle Digital Assistant pricing analysis for any Fusion-embedded chatbot routing through Select AI.

Related reading