What's New - Q4 2025 through June 2026

Material changes that affect architecture, cost, or risk decisions. Curated, not a press-release dump.

Major releases timeline

Practical implications for architects

Service Map

Every Oracle AI service worth knowing, in one diagram. Use this to orient before you go deep.

Reading the map

The top band is where you start if you are buying outcomes, pick a Fusion agent, hook up APEX RAG, or use a managed Gen AI Agent. The middle band is where you start if you are building, pick a model, write prompts, expose endpoints. The bottom two bands are where the data and compute live; you do not get to ignore them, because they drive cost and latency.

OCI Generative AI Service GA

Oracle's managed foundation-model platform. Use hosted models, import compatible models, build Enterprise AI Agents, apply guardrails, or rent a dedicated AI cluster. Region-restricted, IAM-governed, OCI-billed.

OCI Generative AI Agents GA ENTERPRISE AI AGENTS

Oracle's managed agent runtime. Build RAG agents, tool-using agents, and Responses API applications with vector stores, knowledge bases, hosted tools, session state, and governance hooks.

Foundation Model Catalog

Model families and model-delivery paths to understand on OCI Generative AI as of June 24, 2026. Always confirm exact model IDs and region availability in the OCI Console before implementation.

Models by family

Cohere: Oracle's strategic partner

Meta: open weights, broad fit

xAI Grok

NVIDIA

Other delivery paths: Gemini, gpt-oss, and imported compatible models

Choosing a model: quick heuristic

Embeddings & Rerank

The unglamorous half of RAG. Get embeddings wrong and the LLM has no chance.

What embeddings actually are

An embedding is a fixed-length vector of floats that represents semantic content. Two passages that mean similar things produce vectors that point in similar directions. Vector search finds the K nearest neighbors of a query vector and returns the passages they came from. That is retrieval. The LLM then writes an answer from those passages. That is the generation.

Cohere Embed 4: the default on OCI

Chunking discipline (the part teams skip)

• Chunk by structure first, length second. Split on headings, paragraphs, table rows, not arbitrary character counts.
• Aim for ~300-500 tokens per chunk. Smaller chunks improve precision; larger improve context.
• Overlap by 10-15%. Prevents losing the cross-boundary sentence.
• Carry metadata. Source URI, page number, last modified, owning department. You will need this for filtering and citations.
• Re-embed on policy change. Switching embedding models or dimensions means re-embedding the entire corpus. Plan version, cost, and rollback upfront.

Two-stage retrieval (the pattern that wins)

Hybrid search: keyword + vector

Vector search misses queries like "Form 1099-B" because the model treats it as similar to many other tax forms. Keyword search nails it. Hybrid combines both with a weighted score. Oracle AI Database 26ai's Unified Hybrid Vector Search supports this natively in one SQL. OCI OpenSearch supports it as well via score blending.

Enterprise AI Governance & Guardrails

The platform controls around models and agents: guardrails, private endpoints, API keys, IAM, audit, and network isolation.

What Oracle provides natively

• Guardrails for OCI Generative AI: prompt and response checks for unsafe content, prompt injection, and other policy violations.
• On-demand guardrail evaluation: call guardrails directly around a model request, or compose guardrails into your agent path.
• Dedicated endpoint guardrails: inform/block behavior for dedicated AI cluster endpoints where supported.
• PII detection via OCI Language Service, still useful as a deterministic pre/post filter when you need explicit PII categories.
• Agent-level governance: citations, tool schemas, human-in-loop approvals, and session limits.
• Private networking controls: private endpoints, service gateways, IAM policies, resource principals, and zero-trust network patterns where available.

What still belongs in your architecture

Layered defense pattern

Oracle AI Vector Search GA 26ai · Jan 2026

Vectors as a first-class datatype, inside Oracle Database, indexed by HNSW or IVF, and joinable with relational, JSON, graph, and spatial in a single SQL statement.

Select AI GA

Natural language to SQL inside the database. PL/SQL package, four modes, multiple LLM providers, RAG-capable. Available in Autonomous and on-prem 26ai.

The four modes

Provider integration

Select AI is provider-pluggable. You create an AI profile that names a provider (OpenAI, Cohere, Azure OpenAI, OCI Generative AI) and credentials, then attach it to a session. Switching providers is a config change, not a code change. Credentials live in Vault.

Where it fits in an enterprise

Risks specific to Select AI

• Schema disclosure. The LLM sees your table and column names. If those reveal sensitive structure, scope it via grants and avoid passing schemas with regulated-data hints in their names.
• Wrong SQL that runs. The model may produce SQL that returns wrong numbers without erroring. Prefer showsql for non-trivial questions and let a human approve.
• Cost surprises. A natural-language question can produce a SQL that table-scans a fact table. Add query timeouts and resource manager plans.
• Cross-database queries. Don't expect the model to understand database links or sharded topologies without explicit metadata coaching.

In-Database ONNX Embeddings

Load an embedding model into Oracle AI Database 26ai. Generate vectors with a SQL function. No network call, no API key, no per-character cost.

The pattern

Most embedding pipelines call out to a hosted model (OCI Gen AI, OpenAI). That introduces latency, cost per call, and a data-leak surface. In-DB ONNX inverts the dependency: you load the embedding model into the DB once, then call VECTOR_EMBEDDING(text USING model_name) as a function in any SQL. Embeddings happen on the DB server.

Why architects care

• No egress. Embedding data never leaves the DB box. Critical for regulated content.
• No per-call cost. Pay for CPU you already own, not per million characters.
• Lower latency on bulk re-embed. Eliminate network round-trip per chunk.
• Simpler ops. No external service dependency in the embedding pipeline.

Trade-offs

Where it slots in

Oracle AI Database Private Agent Factory 26ai

A no-code/private agent factory for enterprise data. Use it when business users or engineers need knowledge agents grounded in approved repositories, files, web sources, and Oracle Database data without exposing the workflow through a general-purpose SaaS chatbot layer.

Reference architecture

Use it when

• The corpus is private enterprise content: database rows, internal sites, file shares, SharePoint, Google Drive, or uploaded documents.
• You need no-code agent creation for business users while preserving engineered controls around approved sources and model management.
• You need explainable retrieval over documents and vectors without standing up a separate vector DB.

Do not use it when

• The agent is primarily a Fusion process agent covered by Fusion Agentic Apps or AI Agent Studio.
• The corpus is mostly non-Oracle documents in object stores and a managed OCI Generative AI Agent would ship faster.
• You need a consumer-grade assistant UX with broad channels, analytics, and bot lifecycle tools; evaluate Oracle Digital Assistant or app-layer tooling.

Oracle Private AI Services Container 26ai

A lightweight containerized web service for Oracle AI Database 26ai that offloads expensive vector work outside the database: embedding generation and HNSW vector-index creation.

Architecture decision

Two services in the container

Risks

• Model freshness. You manage embedding model updates; stale embeddings quietly degrade retrieval quality.
• Capacity sizing. Offloaded vector work shifts latency and throughput onto your container hosts.
• Patch ownership. Treat the container like production infrastructure, not a demo appliance.
• Endpoint security. Protect the container endpoint and API key; it can be invoked by database jobs or REST clients.
• Audit consistency. Log embedding/index jobs, model versions, container version, target table/index, and caller.

OCI Vision GA

Pretrained and custom-trainable image analysis. Object detection, classification, OCR, document image understanding. API + Console + SDK.

Capabilities

Indicative pricing (verify on the OCI pricing page)

Pretrained image analysis is in the low-cents-per-thousand-images range. Custom model training is hourly per GPU-hour. Always check current numbers before committing.

When to use Vision vs Document Understanding

Risks

• Custom-trained models drift as products and packaging change. Retrain quarterly or on accuracy degradation alarms.
• OCR accuracy degrades on low-quality scans. Pre-process (deskew, contrast) before sending.
• Face detection has compliance implications. Document the legal basis before deploying.

OCI Language GA

NLU primitives for text: sentiment, entity recognition, PII detection, key phrase extraction, language detection, classification, translation.

Capabilities

Language coverage

Most analytical features (sentiment, NER) cover English, Spanish, French, German, Portuguese, Italian out of the box. Coverage varies by feature, check the docs per service. For broader language coverage, pair with an LLM via Gen AI.

Where Language slots into Gen AI

The pattern that works: use Language as cheap pre/post filters around expensive LLM calls. Detect language to pick the right system prompt, scrub PII before sending to the model, classify intent to skip the LLM when a deterministic answer exists. This drops Gen AI cost by 30-60% on a typical customer-support workload.

OCI Speech GA

Speech-to-text (ASR) for audio files and streams. Multiple languages, speaker diarization, SRT/VTT output, profanity handling.

Capabilities

• Batch transcription of audio/video files in Object Storage.
• Real-time streaming for low-latency captioning use cases.
• Speaker diarization ("who spoke when") for call recordings and meetings.
• Normalization of times, addresses, numbers, URLs in the output text.
• Profanity filter: remove, mask, or tag.
• SRT/VTT subtitle output for video.
• Custom vocabulary for domain words the base model mis-hears.

Common architectures

Risks

• Background noise drops accuracy. Pre-process or use a dedicated noise-suppression step before transcription.
• Diarization struggles with overlapping speakers. Document accuracy expectations to stakeholders.
• Audio data residency matters more than most teams think, keep buckets in the right region.
• Real-time streaming has stricter quotas; plan capacity before peak loads (e.g. live earnings calls).

xAI Voice on OCI Generative AI May 2026

Text-to-speech through the OCI Generative AI model layer. Treat it as output generation for voice agents, training narration, call-center assist, and accessibility workflows.

Reference pattern: voice agent response

Use it when

• You need voice output from an Oracle-hosted Gen AI workflow without procuring a separate TTS vendor.
• You are building an internal assistant, training-content generator, or call-center agent response path.
• You can tolerate region/model availability checks and quality testing before launch.

Risks

• Latency. Voice adds another model call after generation. Stream audio where possible.
• Unsafe audio. Guardrail text before synthesizing. Audio moderation after generation is harder.
• Voice consistency. Pin the voice/model choice and test regression on every catalog update.
• Cost. Cache repeated announcements and training clips instead of regenerating.

OCI Document Understanding GA Generative extraction 2026

Extract text, tables, key-value pairs, signatures, and classifications from PDFs and document images. 2026 update added generative extraction for context-aware parsing.

Capabilities

Pricing structure

Charged per transaction (a page or a document, depending on the operation). First 5,000 transactions per month are free: useful for low-volume pilots and Always-Free tier exploration.

Reference pipeline

Risks

• Custom KV models need labeled data, budget annotation time honestly.
• Generative extraction is more accurate but slower and more expensive than classic OCR-only. Mix modes based on document type.
• Tables with merged cells or nested headers still cause issues, sample your hardest documents in PoC.

OCI Anomaly Detection GA

Multivariate time-series anomaly detection using Oracle Labs' MSET2 algorithm. Trained on your historical normal-operation data; scores new observations as anomalous or not.

Why this exists

Most enterprise anomaly problems aren't univariate (a single sensor spike). They're multivariate, "this combination of pressure, temperature, vibration, and current is unusual together even though no single value is out of spec." MSET2 (Multivariate State Estimation Technique) was developed at Oracle Labs for nuclear plant monitoring; it generalizes to manufacturing, fleet telemetry, and IT ops.

How it works

• Train on a window of historical "normal" data, sensor values, KPI series, whatever is multivariate and time-aligned.
• Score new observations, returns an anomaly score per timestamp and per signal.
• Explain: identifies which signals are contributing to the anomaly.

Where it fits

Risks

• Concept drift. What was normal six months ago isn't now. Retrain on a rolling window.
• Cold start. Needs enough clean historical normal data, typically weeks to months.
• False positives. Tune detection sensitivity per use case; pair with operator runbook.
• Not a forecasting service. Use OCI Forecasting if you need next-value prediction.

OCI Forecasting GA

AutoML for univariate and multivariate time-series. Pick a target series, optional exogenous regressors, and a horizon; the service auto-selects and trains a model.

What it gives you

• Point forecasts + prediction intervals.
• Auto algorithm selection across classical (ARIMA, ETS) and ML (Prophet-style, gradient boosted) approaches.
• Holiday and seasonality handling out of the box.
• Multi-horizon forecasts.
• Explanation of which features drive a forecast.

Common use cases

Risks

• Garbage in, garbage out, handle missing values and outliers upstream.
• Forecasts are only as good as the regressors you provide; promos, holidays, pricing must be fed in if they drive the series.
• Auto model selection isn't auto governance, log the chosen model + features per retrain cycle for audit.

Oracle AI Data Platform 2026

The governed data plane for AI teams. Use it to organize data products, catalogs, connections, metadata, workspaces, notebooks, and pipelines so agents and models do not each invent their own data access layer.

Architecture role

When to use it

Risks

• Governance theater. Catalog entries without owners, refresh SLAs, and access rules do not help agents.
• Pipeline duplication. Decide which datasets, catalogs, and metadata flows are authoritative and version them like APIs.
• Latency. A shared data plane is not always the fastest path for hot transactional reads; keep OLTP queries close to the source DB.
• Team boundaries. Data product owners must participate in AI design, or agents will be grounded on misunderstood data.

OCI Data Science GA

JupyterLab notebooks, MLOps pipelines, model catalog, model deployment, jobs, monitoring. Built on conda environments and an Operator pattern.

Building blocks

Where it fits in the AI stack

Data Science is the "build your own" layer. If your problem is solved by Cohere or Llama as-is, you don't need Data Science, use Gen AI Service. If you need a custom model (classical ML, fine-tuned open-weight LLM, vision model, time-series), you do. Many enterprises end up using Data Science only for the 10-20% of use cases that don't fit a managed AI service, the rest go through Gen AI Service or AI Quick Actions.

Risks and ops realities

• Idle notebook spend. Notebook sessions running on a GPU cost real money even when idle. Auto-stop policies are essential.
• Environment sprawl. Teams customize conda environments; reproducibility erodes. Pin environments per project.
• Model-to-production gap. Notebook code rarely runs cleanly as a Job. Budget time for the productionization step every project.
• Compliance for training data. Training datasets often contain PII. Treat them with the same controls as the source-of-record DB.

AI Quick Actions GA Llama 4 + gpt-oss 2026

No-code foundation-model deployment and fine-tuning. Pick a model from a catalog, click Deploy, get an endpoint. Or pick a model, point at training data, click Fine-tune.

What's in the model catalog (June 2026)

• Meta Llama 4: Scout, Maverick.
• Meta Llama 3.x: including Llama 3.2 90B Vision Instruct.
• OpenAI open-weight: gpt-oss-120b, gpt-oss-20b.
• Phi, Falcon, Mistral, Granite, pre-cached, faster cold start.
• Bring-your-own from Hugging Face, direct import.

What it actually saves you

If you've ever deployed an open-weight LLM on raw GPU instances, you've burned days on Docker images, vLLM/TGI tuning, autoscaling, health checks, log shipping. AI Quick Actions does all of that with one click. You give up some flexibility (you can't pick exactly which serving runtime version, for instance) for an order of magnitude faster time-to-endpoint.

When AI Quick Actions vs Gen AI Service

Risks

• Dedicated GPU cost, deployment runs hourly regardless of traffic. Auto-shutdown for dev/test environments.
• Model size vs shape, Llama 4 Maverick won't fit on a single GPU; AQA picks the right shape, but you should understand the floor cost.
• Fine-tuning quality depends on data quality more than algorithm choice. The clicky UI doesn't change that.

Fusion Agentic Applications GA · Mar 2026

22 pre-built AI agents embedded inside Oracle Fusion Cloud. Native to the transactional system. Governed by Fusion roles, approvals, and data. CX expansion in Apr 2026.

What "agentic" means here

Oracle's definition: outcome-driven, proactive, reasoning, and engineered for enterprise execution. Concretely, an agentic app does four things a copilot doesn't: it (1) initiates work without a user prompt, (2) plans across multiple steps and tools, (3) executes within the transactional system using existing roles, and (4) measures outcomes back. The boundary between "agent" and "automation" is fuzzy, but the integration depth is the meaningful difference.

Where the 22 agentic apps sit (representative: Oracle keeps adding)

Why this changes build-vs-buy math

Before Fusion Agentic Apps, an enterprise wanting a "smart period-close" agent had to (a) buy or build an LLM platform, (b) integrate it with Fusion Financials data, (c) replicate role permissions, (d) wire it into approval workflows, (e) operate it. Now Oracle ships steps (a)-(e) as a configured agentic app under your existing Fusion subscription. The build case has to clear a much higher bar.

Risks

• Change management. Agentic apps execute work, that means human approvers see different workflows. Governance and communications matter more than the tech.
• Configuration drift. Each agentic app has settings. Track them per environment and put them under change control.
• Data quality exposed. Agents reason on Fusion data. If your master data is messy, agents are less useful. Fix data first.
• Vendor coupling. Deeper dependency on Fusion. Be intentional about which agents you adopt and which you keep optionality on.

AI Agent Studio for Fusion Expanded Mar 2026

Build, connect, and orchestrate agents that work alongside Fusion. Includes the Agentic Applications Builder, content intelligence, contextual memory, ROI measurement, and workflow tools. Included with Fusion subscriptions at no extra cost.

What you can build

• Custom agents on top of Fusion data and APIs, using Oracle, partner, or external agents as building blocks.
• Agentic applications (workflows of agents) via the no-code Agentic Applications Builder.
• External integrations: pull in Slack, Teams, Microsoft 365, ServiceNow, and similar.
• ROI dashboards: measure agent impact (time saved, cycle time, decision accuracy).

Studio vs OCI Generative AI Agents: what's the difference?

They are complementary, not redundant. A bank could use OCI Generative AI Agents for an internal policy chatbot over a SharePoint-style document corpus, and AI Agent Studio for finance-close agents that operate inside Fusion ERP.

Oracle Digital Assistant GA

Oracle's conversational-assistant platform for enterprise channels. Use it when the hard problem is bot lifecycle, skills, channel delivery, and human-agent handoff rather than raw model prompting.

ODA vs Enterprise AI Agents

Reference pattern

Risks

• Wrong layer. Do not rebuild RAG plumbing in ODA skills when Enterprise AI Agents already provides it.
• Channel complexity. Each channel has identity, session, and attachment quirks; test the exact deployment channel.
• Handoff design. Human handoff must include transcript, context, model answer, and source citations, not just "transfer to agent."

APEX AI GA 24.2 RAG & AI configs

Low-code AI inside Oracle APEX. AI Assistant for developers and end users, AI-driven data modeling, dynamic-action generative text, AI Configurations + RAG Sources, vector search integration.

What APEX 24.2 adds (relevant to AI)

Provider support

APEX talks to OCI Generative AI, OpenAI, Cohere, and Azure OpenAI through provider configurations. The AI Configuration abstracts the provider, apps don't change when you swap.

Reference pattern: RAG chatbot in APEX in a day

1. Create a VECTOR column on your content table in 23ai/26ai. Populate it (in-DB ONNX or Cohere via DBMS_VECTOR_CHAIN).
2. Create a RAG Source pointing to that table.
3. Create an AI Configuration with a system prompt + that RAG Source + your preferred provider.
4. On any page, add the "Show AI Assistant" dynamic action bound to that configuration.
5. Ship.

Risks

• Provider credentials live in APEX Web Credentials, protect them like any other secret.
• Embed cost is real if you store all your content in the DB and embed via OCI Gen AI. Use in-DB ONNX where possible.
• Audit trail of LLM calls isn't automatic, log via APEX_AI calls to a log table for compliance.

MySQL HeatWave GenAI GA

In-database LLMs, automated vector store, lakehouse access, and natural-language chat, all inside MySQL HeatWave. Multilingual, JavaScript-callable, VLM-enhanced PDF parsing as of MySQL 9.4.2.

Components

When to choose HeatWave GenAI over OCI Gen AI + 23ai

• You already run MySQL HeatWave for analytics.
• Your source data is heterogeneous (MySQL + Object Storage + S3) and you want lakehouse ingestion.
• You want LLM inference inside the database (no egress to an external service).
• You don't need Oracle Database features (PL/SQL, VPD, RAC).

Risks

• Sizing, in-DB LLM inference is GPU-intensive on the HeatWave node. Pick shapes deliberately.
• Available model family inside HeatWave is narrower than OCI Gen AI's catalog.
• HeatWave is OCI-first; some features are not available on AWS or Azure deployments of HeatWave.

AI Infrastructure: GPU shapes & networking

What you actually rent when you need raw inference or training capacity. NVIDIA H100 / H200 / B200 / GB200 NVL72, RDMA cluster networks, dedicated regions, sovereign deployments.

Shape family (illustrative: confirm in OCI Compute docs)

Cluster networks & RDMA

For multi-node training, GPUs talk to each other faster than they talk to anything else. OCI Cluster Networks use RDMA over Converged Ethernet (RoCE) with very low latency and high bandwidth between bare-metal GPU nodes. If you're training a model that doesn't fit on one node, this is the lever that determines wall-clock time.

Sovereignty and region matrix

Architecture Patterns

Five reference patterns that cover most enterprise Oracle AI projects. Each names the services, the data flow, and the failure modes.

Pattern 1: Internal RAG chatbot over enterprise documents

Pattern 2: In-database RAG inside an APEX app

For Oracle-shop teams that already use APEX, this collapses the stack dramatically. The data, the embeddings, the search, the chat UI, all inside the database and APEX. Provider call out to OCI Gen AI for generation only. Fastest time-to-prod for internal tools.

Pattern 3: Document automation pipeline

Invoices, claims, contracts arrive as PDFs and need to become structured records. Doc Understanding extracts, validation rules check, exceptions route to humans, results land in the system of record. Add a Gen AI step to summarize or classify when needed.

Pattern 4: Fusion-native agentic workflow

For Fusion customers, this is now the default. Pick an agentic app, configure thresholds and approval routing, monitor outcomes. Custom logic goes into AI Agent Studio. Custom data integrations via Fusion REST APIs or OIC. Almost never needs to call OCI Gen AI directly, the agent uses the embedded LLM.

Pattern 5: Custom fine-tuned model deployment

Niche but real. You have labeled data and a use case where a 7B-13B fine-tuned open-weight model beats prompting a frontier model on cost and accuracy. Pipeline: Data Science notebook for preparation → AI Quick Actions or Gen AI dedicated cluster for fine-tuning → Model Deployment endpoint → integrate via your app. Reserve this for cases where you've already proven a managed model doesn't work.

Decision Matrix

Quick answers to the questions that come up in every architecture review.

Pricing & Cost Control

Where the money actually goes, and the levers that move it.

Cost drivers by service

The cost-control checklist

• Token / character budgets per session. Hard cap. Alarm on breach. Kill on hard breach.
• Rerank to top-K. Reduce context size by 60-80% with no quality loss in most RAG.
• Prompt caching. Where the model supports it, cache the system prompt.
• Audio caching. For xAI Voice, store repeated greetings, disclosures, and training clips rather than regenerate.
• Idle-shutdown on every notebook session, every dev Model Deployment, every dev AQA deployment.
• Tag everything. OCI cost tracking by tag is the only way to attribute spend across teams.
• Quotas. Service limits + Resource Manager quotas prevent a runaway agent from consuming a quarter of budget overnight.
• Two-tier deployments. Cheaper model for routing/classification, expensive model only on the path that needs it.
• Egress. Keep callers inside OCI when calling Gen AI at volume. Outbound from OCI adds up.

Indicative cost shape (illustrative, not a quote)

Risks & Gotchas

The honest list. The stuff you want to hear before pilot, not after go-live.

Model and provider risks

RAG quality risks

Security risks

Operational risks

Strategic risks

• Vendor lock-in. Going deep on Fusion Agentic Apps deepens Fusion lock-in. Worth doing where the apps fit, but be explicit about which workloads stay portable.
• Skills. Oracle AI requires SQL/PL/SQL + cloud + AI knowledge. The unicorn engineer who has all three is rare. Plan training; pair Apps DBAs with data scientists.
• Pace of change. Oracle ships new models monthly. Architectures that assume model stability age fast. Build for swappability.
• Realistic accuracy expectations. 95% accurate is great in lab, terrible if a 5% wrong tax filing creates regulatory exposure. Match accuracy expectation to consequence.

OCI vs AWS vs Azure vs GCP: AI services

Honest four-way side-by-side as of June 24, 2026. Not Oracle marketing, not anyone's marketing. Names and prices move monthly, so verify in each console before you commit.

Generative AI platforms

Model availability, side by side (June 2026)

✓ = first-party / managed · ~ = partial, region-limited, or recently added · ✗ = not native (use direct API or a gateway). Always confirm exact model IDs and regions in the console.

Pricing, normalized (representative, June 2026)

Sources: Oracle OCI Generative AI pricing page; Anthropic / AWS Bedrock pricing; Microsoft Foundry pricing; Google Vertex / Gemini pricing; third-party aggregators (June 2026). Prices change without notice.

Agents & RAG platforms, in depth

Sovereignty & governance

When each wins

Quick alignment (informal)

Sources used for this June 2026 refresh

Primary Oracle docs and blogs the content was anchored to, plus the competitive sources used for the OCI vs AWS vs Azure comparison. Verify the latest before locking in commitments.

June 2026 update (new this refresh)

• What's New in Oracle AI, June 2026 edition (published June 11, 2026)
• Cohere Rerank 4 on OCI (on-demand + dedicated)
• NVIDIA Nemotron 3 Ultra imported-model deployment
• Model Import: Qwen, Gemma, gpt-oss on B200 (release notes)
• OCI Enterprise AI in UAE Central (Abu Dhabi) release notes
• OCI Resource Analytics announcement
• Hybrid RAG in Oracle 26ai as an MCP tool

OCI vs AWS vs Azure vs GCP comparison sources

• OCI Generative AI pricing
• Amazon Bedrock AgentCore GA (Oct 2025)
• Bedrock AgentCore Managed Harness GA (Apr 2026)
• Opening the AWS European Sovereign Cloud (Jan 2026)
• AWS European Sovereign Cloud launch press release
• GPT-5.5 in Microsoft Foundry
• Microsoft Foundry model router (Quality/Cost/Balanced)
• GPT-5 family powers Foundry Agent Service
• Azure AI Foundry renamed Microsoft Foundry (Ignite 2025)
• Anthropic Claude API pricing (Opus 4.8 / Sonnet 4.6 / Haiku 4.5)
• Gemini Enterprise Agent Platform (formerly Vertex AI)
• Introducing Gemini 3 Flash (Google blog)
• Gemini / Vertex AI generative AI pricing
• Run Gemini air-gapped on-prem with Google Distributed Cloud
• Google Model Armor (AI firewall)

OCI Generative AI

• Generative AI documentation
• OCI Generative AI overview
• Generative AI release notes
• OCI Enterprise AI overview
• OCI Enterprise AI GA announcement
• What's New in AI, May 2026 (Oracle blog)
• OCI Gen AI in US Classified Cloud
• Enterprise AI, comparing models on OCI

OCI Enterprise AI Agents and Governance

• Gen AI Agents service overview
• Enterprise AI platform: models, agents, governance
• Announcing the OCI Gen AI Agents RAG service
• AI-powered enterprise search with Gen AI Agents + 23ai + OpenSearch

Oracle AI Database 26ai / 23ai Vector Search

• Introducing Oracle AI Database 26ai
• 26ai release updates (AI Vector Search)
• Oracle AI Vector Search product page
• AI Vector Search key features
• 26ai on-prem Linux x86-64 availability
• Oracle AI Database 26ai AI, ML, and Analytics docs bookshelf
• Oracle AI Database Private Agent Factory User's Guide
• Oracle Private AI Services Container User's Guide
• HNSW and IVF in 26ai (community deep dive)

Oracle AI Data Platform

• Introduction to Oracle AI Data Platform
• Oracle AI Data Platform documentation
• IAM policies for Oracle AI Data Platform Workbench

Select AI

• Autonomous AI Database Select AI overview
• Select AI User's Guide
• Use Select AI for Natural Language Interaction

Fusion Agentic Apps & AI Agent Studio

• Oracle introduces Fusion Agentic Applications (Mar 2026)
• AI Agent Studio expansion (Mar 2026)
• Fusion Agentic Apps for CX (Apr 2026)
• AI Agents for Fusion Applications

APEX AI

• APEX 24.2 GA announcement
• AI Configurations and RAG Sources in APEX
• AI-Driven Data Modeling in APEX 24.2
• Including Generative AI in APEX apps (docs)

Oracle Digital Assistant

• Oracle Digital Assistant documentation
• Oracle Digital Assistant features
• Deploy an Oracle Digital Assistant chatbot powered by OCI Generative AI Agents

MySQL HeatWave GenAI

• MySQL HeatWave GenAI features
• HeatWave GenAI overview (docs)
• New features in HeatWave GenAI

OCI Data Science & AI Quick Actions

• AI Quick Actions overview
• Introducing AI Quick Actions
• Deploy Llama 4 with AI Quick Actions
• OpenAI gpt-oss models in AI Quick Actions

OCI AI Services

• OCI AI Services hub
• OCI Vision
• OCI Language pricing
• OCI Speech pricing
• OCI Document Understanding pricing

Infrastructure / NVIDIA partnership

• Oracle-NVIDIA at GTC 2026
• Oracle expands AI infra options for US Gov