Snowflake Cortex AI
by Snowflake Inc.
Run LLMs, search and agents inside Snowflake, next to your governed data.
Snowflake Cortex AI is Snowflake's managed generative-AI service: SQL-callable LLM functions, hybrid search, text-to-SQL and agents that run on Anthropic, OpenAI, Meta and Mistral models hosted inside Snowflake's security boundary. It lets data teams already on Snowflake add AI to governed data without moving it to a separate model provider or building their own infrastructure.
Snowflake Cortex AI is the generative-AI layer of Snowflake's data platform: a fully managed set of services that run large language models inside Snowflake's security boundary, next to the governed tables, documents, images and audio a company already stores there. It was announced in private preview at Snowday on November 1, 2023, initially offering Llama 2 and serverless LLM functions callable from SQL or Python. Cortex AI Functions (formerly AISQL) now expose SQL functions such as AI_COMPLETE, AI_CLASSIFY, AI_FILTER, AI_AGG, AI_EXTRACT, AI_PARSE_DOCUMENT, AI_TRANSCRIBE and AI_REDACT, backed by Snowflake-hosted models from Anthropic, OpenAI, Meta, Mistral AI and others. Cortex Search is a hybrid vector-plus-keyword engine with semantic reranking for retrieval-augmented generation; Cortex Analyst turns natural-language questions into SQL through Semantic Views; and Cortex Agents orchestrate both, plus custom tools and MCP connectors, through an agent:run REST API. At Snowflake Summit 2026 the knowledge-worker agent Snowflake Intelligence was renamed Snowflake CoWork and the Cortex Code coding agent became Snowflake CoCo; Futurum reported CoCo at more than 9,100 accounts and CoWork at 5,800 in fiscal Q2 2027. Everything is consumption-billed: AI features draw AI Credits, $2.00 each on demand, per million tokens, per 1,000 pages or per GB of indexed data, on top of the warehouse compute that issues the call. Snowflake states that customer inputs and outputs are not used to train models offered to other customers, and access runs through the same role-based privileges as the rest of the account. Its closest rival is Databricks, which VentureBeat flagged at launch; for buyers not already on Snowflake, a direct model API is usually simpler.
The data platform or analytics engineering lead at a company already standardised on Snowflake, because Cortex adds LLM, search and agent capabilities without a new vendor, a data copy or a separate security review.
Run LLM classification, extraction, search and text-to-SQL directly on governed Snowflake data, billed and permissioned like any other Snowflake workload.
At a Glance
- Category
- Data & Analytics
- Pricing
- Usage-based
- Target Market
- CIOs, CTOs, Chief Data Officers, Data Engineers, Data Analysts, Enterprise Developers
- Deployment
- Cloud-only, Multi-cloud
- Founded
- 2012
- Team Size
- 500+
- Customers
- CoCo 9,100+ accounts and CoWork 5,800 accounts as of fiscal Q2 2027 (Futurum); Snowflake overall reported 12,062 customers as of July 31, 2025
Key Features
- ✓Cortex AI Functions
SQL functions such as AI_COMPLETE, AI_CLASSIFY, AI_FILTER, AI_AGG, AI_EXTRACT and AI_SENTIMENT apply LLMs to text, images and audio inside ordinary queries, so analysts add AI to existing pipelines without Python or moving data.
- ✓Multi-vendor model catalog
AI_COMPLETE exposes Snowflake-hosted models including Claude Opus 5 and Sonnet 5, OpenAI GPT-5.x, Llama 3.x, Mistral Large 3, Qwen3 and Gemini 3.1 Pro (public preview), so teams can pick a model per task and cost point.
- ✓Cortex Search
Managed hybrid vector-plus-keyword search with semantic reranking over text columns, refreshed automatically on a configurable TARGET_LAG, giving RAG apps and agents fresh retrieval governed by Snowflake privileges.
- ✓Cortex Analyst
Text-to-SQL over structured data driven by Semantic Views that encode business definitions and metrics, exposed through a REST API for Streamlit, Slack, Teams or custom chat front ends.
- ✓Cortex Agents
Fully managed agent runtime that plans a request, calls Cortex Analyst, Cortex Search, code execution, custom stored procedures or MCP connectors such as Jira and Salesforce, then reflects and responds via the agent:run API.
- ✓Document, audio and PII processing
AI_PARSE_DOCUMENT (OCR or layout mode), AI_TRANSCRIBE and AI_REDACT turn PDFs, recordings and PII-laden text into queryable, redacted columns without exporting files to an outside service.
- ✓Snowflake CoWork and CoCo
Packaged agents built on Cortex: CoWork answers business questions with cited Deep Research and acts in Slack, Gmail, Jira and Salesforce, while CoCo is a data-native coding agent for pipelines, automations and AI apps.
- ✓AI cost telemetry
CORTEX_FUNCTIONS_USAGE_HISTORY and CORTEX_FUNCTIONS_QUERY_USAGE_HISTORY report credit and token consumption per call and per query, so FinOps teams can attribute AI spend by model and workload.
- ✓Cross-region inference control
The CORTEX_ENABLED_CROSS_REGION parameter decides whether prompts may be processed in another region or cloud, from DISABLED through geography-scoped values to ANY_REGION, trading model availability against residency.
Capabilities
Use Cases
- •Tagging tickets and reviews in SQL
Analysts run AI_CLASSIFY and AI_SENTIMENT over large volumes of support tickets or product reviews in a scheduled query, producing tagged tables for dashboards without exporting data to an external API.
- •Governed self-serve analytics
Business users ask questions in plain language and Cortex Analyst generates SQL against curated Semantic Views, so answers use the same metric definitions that finance and operations reporting already rely on.
- •RAG over internal documents
Teams index contracts, policies or product documentation with Cortex Search and ground a Cortex Agent's answers in the retrieved passages, keeping the source text inside the Snowflake boundary.
- •Document field extraction
AI_PARSE_DOCUMENT and AI_EXTRACT pull fields from invoices, forms or PDFs staged in Snowflake into structured columns that downstream pipelines can join, validate and audit.
- •Unifying large content catalogues
Snowflake cites Booking.com unifying 31M travel listings and powering 175K destinations with Cortex AI, a pattern for normalising large, messy catalogues of multilingual content at warehouse scale.
Ideal For
Best For
- ✓Organisations whose governed data already lives in Snowflake and that want LLM features without standing up a separate AI stack
- ✓SQL-first analytics teams adding classification, extraction, sentiment or summarisation to existing scheduled pipelines
- ✓Regulated enterprises that need inference inside an audited boundary (Snowflake lists SOC 2 Type II, ISO 27001, FedRAMP High, HITRUST CSF and PCI DSS)
- ✓Building internal chat-with-your-data agents that combine structured tables (Cortex Analyst) with documents (Cortex Search)
- ✓Teams that want several frontier model vendors under one bill, one access-control model and one usage-history view
Not Ideal For
- ✗Companies not already on Snowflake: Cortex's value is proximity to Snowflake data, and adopting a warehouse just to reach an LLM API is backwards when Bedrock, Azure or Vertex offer those models directly
- ✗Teams with strict in-region processing requirements that will not configure it: organisations created after March 9, 2026 default to cross-region inference (ANY_REGION), which transiently sends prompts and responses to other regions
- ✗High-volume, cost-sensitive workloads without FinOps discipline: AI Credits are billed on top of the calling warehouse, and the platform-credit capacity discount does not apply to AI Credits
Integrations
Deployment
Market & Ratings
CoCo 9,100+ accounts and CoWork 5,800 accounts as of fiscal Q2 2027 (Futurum); Snowflake overall reported 12,062 customers as of July 31, 2025
Market Analysis
Pros
- ✓Data stays inside the Snowflake security boundary, and Snowflake states customer inputs and outputs are not used to train models offered to other customers
- ✓Broad, current model catalog (Claude Opus 5, GPT-5.x, Llama, Mistral Large 3, Gemini 3.1 Pro preview) reachable from one SQL function
- ✓Reuses existing Snowflake RBAC, SSO, compliance attestations and billing instead of onboarding a new AI vendor
- ✓Per-model token rates are published in the Service Consumption Table, and usage-history views make AI spend auditable per query
- ✓Fast adoption signal: Futurum reports CoCo at 9,100+ accounts and CoWork at 5,800 by fiscal Q2 2027
Cons
- ✗Two meters at once: AI Credits are charged per token while the calling warehouse keeps billing, and PeerSpot reviewers of Snowflake already complain that costs 'can increase quickly'
- ✗Agent security is still maturing: PromptArmor showed the Cortex Code CLI could be prompt-injected into escaping its sandbox and running malware (fixed in v1.0.25 on Feb 28, 2026), drawing a 269-point Hacker News thread
- ✗Text-to-SQL quality depends on building and maintaining Semantic Views, and Snowflake's own docs note Cortex Analyst struggles with long multi-turn conversations and cannot see previous query results
- ✗Model and region availability varies, several models are preview-only, and newer accounts default to cross-region inference that moves prompts out of region
- ✗Frequent renaming (AISQL to AI Functions, Snowflake Intelligence to CoWork, Cortex Code to CoCo) makes documentation, training and contract terms a moving target
Pricing
On-demand AI Credits
$2.00 per AI Credit (global); $2.20 regional
- ✓Billed monthly in arrears
- ✓AI_COMPLETE metered per million input/output tokens by model, e.g. claude-sonnet-4-5 at 1.80/9.00 AI Credits
- ✓Cortex Search at 6.3 AI Credits per GB/month of indexed data
- ✓Calling warehouse compute billed separately
Capacity (pre-paid)
From $1.88 per AI Credit (global, $40M+ ACV)
- ✓Tiered by annual contract value: $1.96 from $1.2M ACV, $1.94 from $5M, $1.92 from $10M
- ✓Platform-credit discount does not apply to AI Credits
- ✓Provisioned Throughput reservations available for some AI features
30-day trial
$0
- ✓30 days or until the free usage balance is used
- ✓Trial accounts without a payment method limited to roughly ten credits per day of Cortex AI Functions
Cortex AI is billed in AI Credits, separate from Snowflake platform credits: $2.00 per AI Credit on demand (global; $2.20 regional), falling to $1.88 on capacity contracts above $40M ACV, per the Service Consumption Table effective September 16, 2026. AI_COMPLETE is metered per million input/output tokens by model (claude-sonnet-4-5 at 1.80/9.00 AI Credits, llama3.1-8b at 0.132/0.132), Cortex Search at 6.3 AI Credits per GB/month indexed, AI_PARSE_DOCUMENT layout mode at 3.66 AI Credits per 1,000 pages, and the Cortex Analyst API at 67 platform credits per 1,000 messages. The calling warehouse keeps billing on top, and Snowflake recommends a warehouse no larger than MEDIUM. The platform-credit capacity discount does not apply to AI Credits. Trial accounts without a payment method are capped at roughly ten credits a day of AI Functions.
Security & Compliance
Connect
Sources
This page was written from 22 sources, 18 on domains other than snowflake.com.
- 1.snowflake.com — cortexvendor
- 2.docs.snowflake.com — aisql
- 3.docs.snowflake.com — aisql regional availability
- 4.docs.snowflake.com — aisql cost
- 5.snowflake.com — CreditConsumptionTable.pdfvendor
- 6.docs.snowflake.com — cortex agents
- 7.docs.snowflake.com — cortex search overview
- 8.docs.snowflake.com — cortex analyst
- 9.docs.snowflake.com — cross region inference
- 10.snowflake.com — snowflake ai trust and safetyvendor
- 11.docs.snowflake.com — intro compliance
- 12.docs.snowflake.com — intro editions
- 13.docs.snowflake.com — admin trial account
- 14.snowflake.com — about snowflakevendor
- 15.venturebeat.com — snowflake unveils cortex a managed service to build llm apps
- 16.constellationr.com — snowflake summit 2026 context custom model training iceberg
- 17.techzine.eu — snowflake cortex code is now called coco and performs autono
- 18.futurumgroup.com — snowflake q2 fy 2027 coco and cowork expand ai consumption
- 19.promptarmor.com — snowflake ai escapes sandbox and executes malware
- 20.hn.algolia.com — 47427017
- 21.peerspot.com — snowflake reviews
- 22.en.wikipedia.org — Snowflake Inc.
Stay Ahead of the Curve
Weekly enterprise AI insights for technology leaders. No spam, no vendor pitches—unsubscribe anytime.
SubscribeRelated Products
Tacnode Context Lake
PostgreSQL-compatible context lake giving AI agents shared, live, transactional state at decision time
Reducto
Agentic document parsing API that turns complex PDFs, scans and spreadsheets into citation-grounded data for AI, now with the 1¢-per-page r-1 model
Euno
AI-native context platform that gives enterprise agents governed, trustworthy data
Prevalent AI
A sovereign enterprise knowledge graph that gives security teams and AI agents context