What is Snowflake Cortex? The AI and ML Layer Inside Snowflake

Snowflake Cortex is Snowflake's managed AI and ML service that runs LLM functions, ML model serving, and AI feature engineering inside the Snowflake compute boundary, with no data movement.

Snowflake Cortex is Snowflake's managed AI and ML service, launched in 2023 and expanded heavily through 2024 to 2026. It runs LLM functions, ML model serving, document AI, and natural-language-to-SQL inside the Snowflake compute boundary, with no data movement outside the customer's Snowflake account.

What Cortex actually is

Cortex is a collection of AI capabilities exposed through SQL-friendly interfaces:

• Cortex LLM Functions: SQL functions for sentiment analysis, summarization, translation, classification, and text completion against rows of data, calling Llama, Mistral, Claude, GPT, or Snowflake Arctic under the hood.
• Cortex Search: vector search service that indexes Snowflake tables and serves k-nearest-neighbor queries with built-in embedding generation.
• Cortex Analyst: a natural-language-to-SQL service grounded against a customer-defined semantic model.
• Document AI: PDF and image extraction with structured field outputs.
• Cortex Fine-Tuning: domain-specific fine-tuning of select base models on the customer's Snowflake data.

The shared property across all of Cortex is data locality. The compute runs inside the customer's Snowflake account boundary, the model invocations are billed against Snowflake's per-token rates, and the data never leaves Snowflake to reach an external API.

What people confuse it with

• "Cortex is Snowflake's version of OpenAI." Closer to true that it is Snowflake's AI service. Cortex exposes multiple model providers (including OpenAI, Anthropic, Meta, Mistral, and Snowflake's own Arctic) through a unified interface.
• "Cortex Analyst is text-to-SQL that just works." Production accuracy depends entirely on the semantic model definition. Teams that skip the semantic model work get the same kinds of hallucinations they got from previous-generation text-to-SQL tools.
• "Cortex replaces our RAG stack." For in-Snowflake retrieval, often yes. For multi-source retrieval that spans Snowflake plus Confluence plus Salesforce, Cortex is part of the stack rather than the whole thing.

When Cortex matters

Cortex matters when:

• The organization runs Snowflake as the primary analytical platform and has Snowflake data the AI features need to reach.
• Data locality requirements rule out external API calls (regulated industries, sensitive PHI / PCI workloads).
• The team is SQL-first and wants AI features that compose with the existing analytical workflow.

When Cortex does not help

Cortex does not help when:

• The organization runs Databricks or BigQuery as the primary platform and Snowflake is a peer rather than the system of record.
• The AI workload is customer-facing interactive and the latency budget is sub-second (Cortex LLM Functions are tuned for analytical, not transactional, latencies).
• The team is code-first and wants the flexibility of direct model API access over SQL-function ergonomics.

How Thinklytics works on Snowflake Cortex

We scope Cortex engagements as adjacent to data foundation work, not standalone. The Cortex value comes from clean semantic models, well-defined certified metrics, and good data hygiene. See Snowflake Cortex consulting for the engagement shape.

Frequently asked questions

What is Snowflake Cortex in one sentence?

Snowflake Cortex is Snowflake's managed AI and ML service that runs LLM functions (summarize, sentiment, translate, complete), ML model serving, document AI, and AI feature engineering inside the Snowflake compute boundary, with no data movement outside the Snowflake account.

What are the main Cortex products?

Cortex LLM Functions (sentiment, summarize, translate, complete, classify), Cortex Search (vector search over Snowflake tables), Cortex Analyst (natural language to SQL with semantic model grounding), Document AI (PDF and image extraction), and Cortex Fine-Tuning (model fine-tuning on Snowflake data).

How is Cortex priced?

Per-token for LLM functions (rates vary by model: Llama 3 70B, Claude 3 Sonnet, Mistral Large 2, GPT-4o, etc.). Cortex Search is a separate compute pool. Most production workloads land at $2,000 to $30,000 per month for Cortex usage on top of base Snowflake compute and storage.

Which LLM models does Cortex support?

As of 2026: Llama 3 family, Mistral family, Snowflake's Arctic models, Anthropic Claude models, and OpenAI models. Available models vary by region. Customers can also fine-tune select models on their own Snowflake data.

Is Cortex a competitor to Databricks Mosaic AI?

Yes, in concept. Both run LLMs and ML serving inside the data platform's compute boundary. Cortex is more turnkey for SQL-first teams (the functions look like UDFs). Mosaic AI is more flexible for code-first ML teams. The cross-vendor comparison usually decides on team skills, not feature parity.

Does Cortex replace OpenAI or Anthropic direct API calls?

For analytical workloads against Snowflake data, often yes. The data-locality argument (no data movement) plus the SQL-function ergonomics make Cortex the cheapest path for in-database LLM use. For interactive customer-facing apps and complex agentic workflows, direct API calls usually still win.

What is Cortex Analyst?

Cortex Analyst is Snowflake's natural-language-to-SQL service, grounded against a semantic model the customer defines. It is positioned against Databricks' Genie, ThoughtSpot Sage, and Snowflake's own legacy text-to-SQL features. Production accuracy depends entirely on the quality of the semantic model definition.

How does Thinklytics work on Snowflake Cortex?

We scope Cortex engagements as adjacent to data foundation work, not standalone. The Cortex value comes from clean semantic models, well-defined certified metrics, and good data hygiene. See Snowflake Cortex consulting for the full engagement shape and Snowflake vs Databricks AI workloads for the cross-vendor comparison.