Snowflake Cortex AI Review (2026): Features, Pricing & Verdict
Snowflake Cortex AI is the integrated suite of generative-AI and machine-learning capabilities built directly into the Snowflake AI Data Cloud — and its entire philosophy is captured in one principle: bring AI to where your data already lives, rather than exporting data to an external model. Instead of shuttling sensitive data out to third-party APIs, you invoke LLM functions, embeddings, semantic search and agents directly inside Snowflake, keeping everything within Snowflake’s security perimeter, governed by the same RBAC, masking policies and audit logs you already run. It has matured dramatically between 2023 and 2026, expanding from a handful of simple LLM functions into a comprehensive, AI-native execution layer. The suite spans several tightly-integrated components: AISQL (Cortex LLM functions) that let analysts call AI models like AI_COMPLETE, AI_CLASSIFY, AI_EXTRACT, AI_SENTIMENT and the context-aware AI_REDACT straight from a SQL SELECT statement — no Python, no ML framework, no data movement; Cortex Analyst, a fully managed text-to-SQL service that turns natural-language business questions into accurate, Snowflake-aware SQL via semantic models; Cortex Search, a managed RAG service blending semantic vector and keyword search for chatbots and document discovery; Cortex Agents (with the no-code Snowflake Intelligence layer on top) for orchestrating multi-step workflows across structured and unstructured data; Document AI, using Snowflake’s proprietary Arctic-TILT model to extract data from invoices, contracts and PDFs; plus Cortex Fine-tuning, Cortex Code and AI Function Studio. Behind the scenes it routes to foundation models from OpenAI, Anthropic, Google, Meta and Mistral, fully serverless with automatic scaling and — importantly — Snowflake never trains those base models on your data. The honest counterweights: token-based AI Credit pricing that is fundamentally different from Snowflake’s familiar warehouse billing and can spike dramatically if unmonitored, continuously-billing serving compute for Cortex Search, meaningful Snowflake lock-in, and a data-warehouse-centric orientation.
- Best for
- Organisations already on Snowflake wanting to run generative AI, RAG, text-to-SQL and agents directly on their governed data — especially SQL-native teams — without moving data or managing GPUs
- Platform
- AI suite inside the Snowflake Data Cloud; AISQL functions, Cortex Analyst, Cortex Search, Cortex Agents, Snowflake Intelligence, Document AI, Fine-tuning, Cortex Code; serverless; via SQL/Python/REST API
- Key differentiator
- AI where your data lives — invoke LLMs, RAG and agents directly in SQL inside Snowflake’s security perimeter, no data export, existing governance preserved
- Pricing
- Consumption via new “AI Credits” (from April 2026); AI functions/agents per M tokens (model-dependent); Cortex Analyst per 1,000 messages; Cortex Search serving + embedding compute; generated SQL uses warehouse credits
- Vendor
- Snowflake — Cortex AI is its native gen-AI/ML framework; routes to OpenAI, Anthropic, Google, Meta, Mistral models; Snowflake does not train base models on customer data
What Is Snowflake Cortex AI?
Cortex AI is Snowflake’s answer to the same question Databricks answers with Mosaic AI: how do you do enterprise AI without ripping your data out of the platform it already lives in? Snowflake’s response is to make its Data Cloud an AI-native execution environment — rather than treating AI as a separate stack you send data to, Cortex integrates intelligence directly into your data workflows, which eliminates the operational, security and compliance headaches that traditionally accompany enterprise AI adoption. The core principle, repeated throughout the product, is “AI where your data lives”: every Cortex capability operates within Snowflake’s governed boundary, respecting existing RBAC controls, dynamic masking, data-residency requirements and audit logging, so proprietary and sensitive data never leaves the security perimeter and — crucially — is never used to train the underlying foundation models. Execution is fully serverless: Snowflake routes requests to foundation models from OpenAI, Anthropic, Google, Meta and Mistral, scaling automatically without you provisioning infrastructure or managing GPU clusters. What makes Cortex distinctive among AI platforms is accessibility across personas: a business analyst can explore data by asking questions in plain English through Cortex Analyst or the no-code Snowflake Intelligence; a SQL-native analyst can run AI functions inside ordinary queries; and a data engineer can build sophisticated RAG applications and multi-step agents — all on the same governed data, without leaving Snowflake. Within this site’s Machine Learning & MLOps category, Cortex AI is the data-warehouse-native counterpart to Databricks Mosaic AI and the hyperscaler platforms — the natural choice for the vast population of enterprises whose data foundation is Snowflake, and this review weighs where that “AI in the warehouse” model excels and where its costs and constraints bite.
Core Features
AISQL and Cortex LLM functions: AI inside your SQL
AISQL is the most accessible and most-used entry point into Cortex, and it’s genuinely clever: it lets anyone who can write SQL invoke AI directly inside a query, with no Python, no ML framework and no data movement. You simply add an AI function as a column in a SELECT statement — for example SELECT AI_SENTIMENT(review_text) FROM product_reviews — and Snowflake runs the model against your data in place. The function library is broad and practical: AI_COMPLETE for text generation, AI_CLASSIFY for categorisation, AI_EXTRACT for pulling structured fields from unstructured text, AI_SENTIMENT, AI_SUMMARIZE, AI_TRANSLATE, AI_EMBED, AI_SIMILARITY, AI_TRANSCRIBE, AI_PARSE_DOCUMENT, and helper functions like AI_COUNT_TOKENS and TRY_COMPLETE that reduce failure cases. A standout is AI_REDACT, which uses an LLM to detect and mask personally identifiable information in unstructured text — and because it understands context rather than relying on regex, it catches names, addresses and sensitive IDs embedded in natural language (support tickets, emails) that pattern-matching would miss. You can analyse text, images and audio alongside your structured data, all powered by the multi-vendor model catalogue. The only prerequisite is the CORTEX_USER role, granted to PUBLIC by default, so most users already have access with no API keys or setup. This SQL-native design is Cortex’s single biggest accessibility advantage: it meets the enormous population of SQL-fluent analysts and engineers exactly where they already work. The realistic caveat is the flip side of that ease — because you can invoke a powerful model on a whole table in one line, costs can escalate fast if you pick an expensive model or run against huge datasets without first checking token counts (more on that in pricing).
Cortex Analyst, Cortex Search and Document AI
Three managed services turn Cortex from a function library into a full applications platform. Cortex Analyst is a fully-managed, LLM-powered text-to-SQL service that lets business users ask questions in natural language and receive accurate answers from structured data without writing SQL — and it’s more sophisticated than a naive converter: it uses semantic models to generate highly accurate SQL, handles joins correctly, aligns to Snowflake syntax, and supports verified queries (pre-approved SQL for common business questions that both improves accuracy and cuts cost by reusing trusted scripts). Available as a REST API, it drops into any application, and it’s the engine behind reliable conversational self-service analytics. Cortex Search is a fully-managed RAG service for unstructured data that combines a hybrid strategy — semantic vector search plus keyword (lexical) matching — so it returns relevant results for fuzzy natural-language queries while still nailing exact matches like a specific item ID, legal term or person’s name that pure vector search can miss; it powers AI chatbots and document discovery across large collections such as millions of support tickets. Document AI rounds it out, using Snowflake’s proprietary Arctic-TILT LLM to extract data from document formats — handling both text-heavy content and graphical elements like logos, handwritten signatures and checkmarks — to automate pipelines for continuously processing invoices, financial statements and PDFs. Together these give teams the three building blocks of modern data AI (structured querying, unstructured retrieval, and document extraction) as governed, managed services inside Snowflake, rather than a stack of external tools to integrate and secure. The realistic note is that the managed convenience comes with service-specific pricing models — Cortex Search in particular carries continuous serving-compute cost — that need understanding before production rollout.
Agents, Snowflake Intelligence and the governed-AI advantage
The top of the Cortex stack is agents and orchestration, and it’s where Snowflake ties everything together. Cortex Agents is the developer platform for building programmatic, multi-step workflows that coordinate across the other services — an agent can, for instance, automatically retrieve customer contracts (Cortex Search), analyse sales data (Cortex Analyst) and generate renewal recommendations (an LLM), all in one orchestrated flow across structured and unstructured data. On top of that sits Snowflake Intelligence, a no-code, business-user interface over the same agent capabilities, adding cross-document analytics and native integrations with tools like Jira and Salesforce so non-technical users can benefit from agentic AI without writing code. The suite is rounded out by Cortex Fine-tuning (adapting foundation models to your domain terminology without leaving Snowflake, which reduces token usage over time), Cortex Code (launched November 2025 — AI-assisted development via Snowsight or the CLI, supporting dbt and Apache Airflow, powered by models like Claude Opus and GPT-5-class), and AI Function Studio (2026 public preview, automating prompt engineering, model selection and benchmarking). But the through-line — and Cortex’s genuine competitive advantage — is governance. Because every capability runs inside Snowflake’s perimeter, you get enterprise AI that respects all existing governance policies, RBAC controls and data-residency requirements by default, and Snowflake does not use your data to train the base models. For regulated enterprises, that “AI without the data leaving the governed boundary” model removes the single biggest blocker to enterprise AI adoption. The realistic caveat is the mirror image: this advantage is fully realised only when your data is in Snowflake, so the more you build on Cortex, the deeper the Snowflake lock-in.
Scored Categories
Pricing
| Component | Model | Notes |
|---|---|---|
| AI Credits (new) | Separate currency | From April 1, 2026, AI features bill in AI Credits, separate from Platform Credits; existing capacity balance and contract terms unchanged, no amendment required |
| AISQL / AI functions | Per million tokens | AI_COMPLETE, AI_CLASSIFY, AI_EXTRACT etc. billed per M tokens; both input and output count; rate varies by model; prompt caching billed at a reduced rate |
| Cortex Analyst | Per 1,000 messages | Standalone API billed per 1,000 messages; via Cortex Agents it’s token-based instead; executing the generated SQL incurs standard warehouse compute |
| Cortex Search | Serving + embedding | Serving compute (credits/GB/month of indexed data) runs continuously while resumed; embedding compute per token on insert/update — cost even when idle |
| Agents / Intelligence / Doc AI | Token / per-page | Agents and Snowflake Intelligence per M tokens (additive across invoked services); AI Parse Doc per 1,000 pages by mode |
COUNT_TOKENS on a sample before processing large tables; pick the cheapest adequate model (a mini model often costs ~10× less than GPT-4/Claude Opus for classification or sentiment and works fine); use verified queries in Cortex Analyst and prompt caching where possible; monitor via CORTEX_FUNCTIONS_USAGE_HISTORY and set scheduled-task alerts to Slack/Teams; and reserve Cortex Search for cases where semantic search is genuinely core. Snowflake notes Cortex can be 10–40× cheaper than third-party APIs when used well — the ROI is excellent with discipline, painful without it. Verify current rates in Snowflake’s Credit Consumption Table.Strengths
- AI runs where your data lives — no data export, no external APIs
- AISQL: invoke powerful AI models directly from SQL, no Python needed
- Governed by default — RBAC, masking, audit, data residency preserved
- Snowflake never trains base models on your data — strong privacy
- Comprehensive suite: functions, Analyst, Search, Agents, Document AI
- Cortex Analyst text-to-SQL with semantic models and verified queries
- Cortex Search hybrid (vector + keyword) RAG for accurate retrieval
- Fully serverless — no infrastructure or GPU cluster management
- Multi-vendor models (OpenAI, Anthropic, Google, Meta, Mistral)
- Accessible to every persona: analysts, engineers, business users
Weaknesses
- Token-based AI Credit pricing is unpredictable and can spike sharply
- A single unmonitored query can burn thousands of credits
- Cortex Search serving compute bills continuously, even when idle
- Both input and output tokens count — easy to under-estimate cost
- Requires disciplined cost monitoring to avoid budget overruns
- Meaningful Snowflake lock-in — value depends on data being in Snowflake
- Data-warehouse-centric; less flexible for app-centric GenAI systems
- New AI Credits billing currency adds a layer of cost complexity
Verdict: 8.3 / 10 — The Governed, SQL-Native AI Layer for the Snowflake Data Cloud
Snowflake Cortex AI earns an 8.3, a strong showing that mirrors its rival Databricks Mosaic AI’s data-platform-native approach from the warehouse side. Its case is compelling and specific: it brings AI to where your data already lives, invokable directly from SQL, governed by default within Snowflake’s security perimeter, with no data movement and no model training on your data — removing the security, compliance and integration blockers that stall enterprise AI. The suite is genuinely comprehensive (AISQL functions, Cortex Analyst text-to-SQL, Cortex Search RAG, Agents, Snowflake Intelligence, Document AI), serverless, and uniquely accessible across analysts, engineers and business users alike. For the large population of Snowflake-native enterprises, it’s an outstanding, low-friction way to operationalise AI on governed data. The score is held back by real, well-documented issues rather than capability gaps: token-based AI Credit pricing that is unpredictable and can spike dramatically without careful monitoring, continuously-billing search serving compute, meaningful Snowflake lock-in, and a warehouse-centric orientation that suits data-heavy analytics better than app-centric GenAI. If Snowflake is your data foundation, Cortex AI is a compelling, best-fit choice — just build cost monitoring in from day one.
Frequently Asked Questions
How is Snowflake Cortex AI different from Databricks Mosaic AI?
They’re the two leading data-platform-native AI suites, and they solve the same problem — do AI where your data lives — from opposite starting points, so the choice usually follows which data platform you already run. Cortex AI is built into the Snowflake Data Cloud and is, at its heart, SQL-native and warehouse-centric: its signature capability is invoking AI directly inside SQL queries (AISQL) on your governed structured and unstructured data, with fully-managed services (Cortex Analyst, Search, Agents, Document AI) layered on top, all serverless and inside Snowflake’s security perimeter. It excels for the vast population of SQL-fluent analysts and for analytics-heavy AI use cases, and its governance-by-default and no-training-on-your-data posture are strong. Databricks Mosaic AI is built on the Databricks Lakehouse (Spark, Delta Lake, Unity Catalog) and is more code-and-engineering-centric: it’s built on open-source MLflow, carries genuine model-training pedigree from the MosaicML acquisition, and offers a comprehensive Agent Framework and an AI Gateway that governs even externally-hosted models. Broadly, Cortex is the more accessible, SQL-first, analytics-oriented choice, while Mosaic AI is the more engineering-oriented, training-and-agent-heavy choice with an open MLflow foundation. Both carry lock-in to their respective platforms and both use complex consumption pricing. The decision rule is straightforward: if your data foundation is Snowflake, Cortex AI removes the most friction; if it’s the Databricks Lakehouse, Mosaic AI does. For most enterprises, existing data-platform commitment outweighs the (real but secondary) capability differences between them.
Is my data safe with Cortex AI, and does Snowflake train models on it?
Data safety is one of Cortex AI’s strongest selling points, and the short answers are: your data stays within Snowflake’s governed boundary, and no — Snowflake does not use your data to train the underlying foundation models. This is central to the “AI where your data lives” design. Because every Cortex capability executes inside Snowflake’s security perimeter rather than sending data to external APIs, all processing respects your existing governance automatically: role-based access control (RBAC), dynamic data masking policies, data-residency requirements and audit logging all continue to apply to AI operations exactly as they do to ordinary queries. You don’t have to build a separate security and compliance regime for AI, and sensitive data never leaves the environment you already trust — which eliminates the operational, security and compliance challenges that traditionally accompany enterprise AI adoption and are often the single biggest blocker to getting AI into production in regulated industries. On model training specifically, Snowflake is explicit that a core advantage of Cortex AI is that it does not use customer data to train the base models; when you call a foundation model from OpenAI, Anthropic, Google, Meta or Mistral through Cortex, Snowflake routes the request but your proprietary and sensitive information remains private and is not used to improve those models. There’s even a context-aware AI_REDACT function to automatically mask PII in unstructured text before further processing. The practical takeaway: for organisations with strict privacy, compliance or data-residency obligations, Cortex AI’s governed-by-default, no-training-on-your-data model is a genuine advantage over sending data to external model APIs — and a major reason enterprises adopt it. As always, verify the specifics against your own compliance requirements and Snowflake’s current documentation.
Why can Cortex AI bills be unpredictable, and how do I keep costs under control?
Cortex AI bills are unpredictable because AI features use token-based consumption pricing (billed in Snowflake’s new AI Credits) that behaves fundamentally differently from the warehouse-based pricing Snowflake users are accustomed to — and that difference catches teams out. With warehouse pricing you reason about compute size and runtime; with token pricing, cost scales with the volume of text processed, and because you can invoke a model against an entire table in a single line of SQL, a seemingly small query can process enormous token volumes. The real-world warnings are sobering: teams have burned significant credits in an afternoon “quick-testing” with an expensive model, and one company paid nearly $5,000 for a single query that processed 1.18 billion records — driven by token costs, not compute. Two structural gotchas compound it: both input and output tokens count toward your bill (output is easy to overlook), and Cortex Search’s serving compute bills continuously while the service is resumed, even when it’s idle. To keep costs controlled: before processing a large table, run COUNT_TOKENS on a sample to see what you’re dealing with (token counts are almost always higher than expected); choose the cheapest model that does the job, since a small/mini model frequently costs around 10× less than GPT-4 or Claude Opus and works fine for classification, sentiment or extraction; use Cortex Analyst’s verified queries and prompt caching to avoid regenerating and reprocessing; monitor spend with views like CORTEX_FUNCTIONS_USAGE_HISTORY and wrap them in scheduled tasks that alert Slack or Teams; and reserve continuously-billing services like Cortex Search for cases where semantic search is genuinely core. Used with this discipline, Snowflake positions Cortex as 10–40× cheaper than third-party APIs — the ROI is excellent, but only if you build monitoring in from the start rather than discovering the bill after the fact.