Snowflake Cortex

Snowflake Cortex is the native AI and machine learning suite built into the Snowflake Data Cloud, developed by Snowflake Inc. It lets data teams run large language model inference, build AI agents, and deploy semantic search directly against data that already lives in Snowflake, without routing that data to external AI services or managing separate GPU infrastructure. The core proposition is security by architecture: LLMs come to your governed data rather than your data leaving to reach the LLMs. Every Cortex operation inherits Snowflake's existing access controls, audit logging, VPC boundaries, and compliance certifications.

The Cortex suite has four distinct service pillars. Cortex LLM Functions (also called Cortex AI SQL) lets analysts call models like Anthropic Claude, Meta Llama 4, and Mistral directly from SQL using functions such as AI_COMPLETE, AI_CLASSIFY, AI_SUMMARIZE, SENTIMENT, TRANSLATE, EXTRACT_ANSWER, and AI_TRANSCRIBE. Cortex Search provides a fully managed hybrid search service that combines semantic vector retrieval with keyword matching across unstructured documents stored in Snowflake. Cortex Analyst converts plain-English questions into SQL and returns prose answers, making structured data accessible to business users who don't write queries. Cortex Agents orchestrates all of the above into a multi-step reasoning loop exposed via a REST API, allowing teams to build chatbots and automated workflows that span both structured tables and document stores. A fifth component, Cortex Code, is a Snowflake-native AI coding assistant with full awareness of the account's schemas and column names, adopted by over 50% of Snowflake customers since its November 2025 launch.

What Snowflake Cortex actually does in May 2026

Cortex reached production readiness on November 4, 2025, when Snowflake simultaneously shipped Cortex Agents and Cortex AI Functions to General Availability at the Snowflake BUILD 2025 developer conference. Before that date, both services had been in preview, limiting adoption to early-access customers. The GA release opened them to all Snowflake accounts and gave enterprises the contract-level SLAs needed to build production workloads on top of them.

The current capability set covers the full AI-on-data workflow: ingest, transform, analyze, and surface. A data engineer processing 10 million customer reviews can add a single SQL column referencing AI_SENTIMENT to an existing dbt model and get sentiment scores on the next pipeline run, with no new credentials to manage and no additional data warehouse to spin up. A team building an internal analytics chatbot uses the Cortex Agents REST API to accept natural language questions, route structured queries to Cortex Analyst for SQL generation, route document questions to Cortex Search for retrieval, synthesize the results, and return a single coherent answer. Multimodal processing extends the same SQL-callable pattern to images and audio: AI_TRANSCRIBE converts audio or video files stored in Snowflake stages directly into text for downstream analysis.

Fine-tuning is also serverless. Teams can fine-tune Llama models on proprietary Snowflake data without provisioning GPU compute or leaving the Snowflake environment. The trained model weights stay inside the account, and serving happens through the same credit-based metering as all other Cortex functions.

"Cortex Code is transforming how we approach agentic analytics at Braze. Its native understanding of our datasets, schemas, and columns means our engineers spend less time wrestling with context and more time getting precise, actionable outputs." - Spencer Burke, SVP of Growth at Braze, Snowflake press release, February 2026

Where Snowflake Cortex sits versus Databricks Mosaic AI and AWS Bedrock

Databricks Mosaic AI is the closest architectural rival because it solves the same problem: AI embedded inside an existing data platform so teams don't build a separate ML stack. The mechanical difference is the abstraction level. Databricks exposes GPU clusters explicitly through the Mosaic AI Training and Serving layers; engineers can inspect node configurations, choose hardware tiers, and run custom PyTorch training jobs with full visibility into compute. Mosaic AI's Agent Framework supports custom model training, reinforcement learning from human feedback, and multi-provider LLM routing via AI Gateway with rate limiting. Snowflake Cortex abstracts all of that away: you write SQL, Snowflake allocates serverless compute invisibly, and there are no GPU configuration choices. Databricks wins for organizations that need to train proprietary models, run rigorous evaluation suites with MLflow integration, or build complex agent architectures that require custom memory and tool registries. Cortex wins for Snowflake shops that want speed-to-production AI without standing up an ML platform. Full comparison at Databricks Mosaic AI.

AWS Bedrock is cloud-agnostic and model-diverse, offering access to 30-plus foundation models from Anthropic, Meta, Mistral, Cohere, Amazon Nova, and others via a unified REST API. Bedrock's architectural contrast with Cortex comes down to data movement. Bedrock operates outside your data warehouse: to use it against Snowflake data, you extract a dataset, send it across a network boundary to the Bedrock endpoint, and receive results back. That boundary crossing requires additional IAM policies, network security controls, and compliance documentation for regulated industries. Cortex eliminates that boundary entirely by running inference inside Snowflake's already-certified security perimeter using the same IAM roles and audit logs your governance team already manages. Bedrock's advantage is breadth: it supports models and agent patterns that haven't yet appeared in the Cortex catalog, and it integrates naturally with AWS Step Functions and Lambda for complex workflow orchestration. For teams whose data spans multiple clouds, Bedrock's cloud-agnostic positioning is architecturally cleaner than Cortex. Compare options at AWS Bedrock, Google Vertex AI, and Azure OpenAI.

What the Cortex workflow reality looks like

For a data engineer already fluent in Snowflake SQL, the Cortex onboarding experience is genuinely low-friction. There is no new platform to learn, no SDK to install for basic functions, and no credentials to manage. A common first project is running AI_SENTIMENT or AI_SUMMARIZE against an existing table of customer records. The query works, the results appear, and the iteration cycle from idea to working prototype is hours rather than weeks.

The friction appears at the cost monitoring layer. Snowflake Cortex now has nine distinct service categories with separate billing views: AI SQL Functions, Cortex Code, Cortex Copilot, Cortex Analyst, Cortex Search, Cortex Agents, Fine-Tuning, Document Processing, and Legacy Functions. These views use inconsistent user identifier fields, making unified cost dashboards non-trivial to build. A practitioner who audited a production Snowflake account found that over 60% of Cortex spend was invisible to the monitoring team because they were only watching the primary AI Functions usage view.

"In one account I audited recently, over 60% of Cortex spend was invisible to the monitoring team." - Vedprakash, data engineer, DataAIChronicles on Medium, March 2026

Cortex Search compounds this because it uses an always-on billing model. Unlike the SQL functions that charge per invocation, Cortex Search services are background processes that run continuously and bill by GB of indexed data per month, regardless of whether any search queries run during that billing period. Teams accustomed to Snowflake's pay-for-compute model are sometimes surprised when a Cortex Search service they spun up for a proof-of-concept accumulates a month of indexing charges.

Snowflake released native spend controls for Cortex AI Functions in March 2026, adding per-user rolling 24-hour limits and team-level budget caps. But robust monitoring still requires stitching together multiple metering views into a unified cost dashboard, something most teams build reactively after an unexpected bill rather than proactively before launch.

Teams building search-powered applications often pair Cortex Search with enterprise knowledge management tools. Products like Glean address similar enterprise search problems for unstructured content, though through a different architectural approach.

Who Snowflake Cortex is built for

The primary user is a data engineer or analytics engineer whose organization has already committed to Snowflake as its data warehouse. Cortex is not a standalone product: there is no separate Cortex subscription, no free tier, and no way to access Cortex without a Snowflake account. If that describes your situation, Cortex is the fastest path to embedding AI into data workflows you already own.

Within that constraint, Cortex is particularly strong for regulated industries. Healthcare, financial services, and insurance organizations operating under HIPAA, SOC 2, or GDPR requirements often have contractual obligations around where data can be processed. Running LLM inference inside a Snowflake account covered by an existing Business Associate Agreement is architecturally simpler than negotiating a new BAA with a separate AI provider for every model vendor in the Bedrock catalog.

Analytics teams building self-service tools for non-technical business users benefit from Cortex Analyst. Rather than building a custom NL-to-SQL layer, they expose the Cortex Analyst API behind a Streamlit or internal chatbot interface and let business users query data in plain English.

What Snowflake Cortex is not

Cortex is not a reason to switch data platforms. If your data lives in Databricks, BigQuery, or Redshift, Cortex is not available to you. The data-residency benefit only applies to data already inside a Snowflake account; there is no agent or connector that reaches into other systems while keeping Cortex's governance properties intact. Cortex Code expanded in February 2026 to reference external systems like AWS Glue and Postgres, but the LLM inference functions themselves still only operate on data staged in Snowflake.

Cortex is not a full ML platform. Fine-tuning support is limited to Llama model variants. There is no custom model architecture training, no reinforcement learning from human feedback, and no equivalent of Databricks' MLflow experiment tracking for systematic model evaluation. Teams whose AI roadmap involves training proprietary models on their own data at scale will outgrow Cortex's managed offering and need to look at Databricks Mosaic AI or a dedicated MLOps platform.

Cortex is also not suitable for teams without cost monitoring discipline. The combination of token-based billing across nine service categories, always-on Cortex Search infrastructure charges, and the absence of native resource monitors until March 2026 has caught teams with unexpected bills. One widely circulated case involved a data team paying approximately $5,000 for a single Cortex Functions query that processed 1.18 billion customer records. The query finished, the credits drained, and no alert had been configured to catch it. That billing dynamic is now partially mitigated by the March 2026 cost controls, but teams should build monitoring dashboards before enabling Cortex broadly across a shared account.