Snowflake Cortex Pricing in 2026: How Much Does It Actually Cost

A single Snowflake query once cost a team close to $5,000. Not a month of usage, one query, an unbounded AI function fired across 1.18 billion records before anyone thought to add a LIMIT. No one had been reckless. They simply did not know how Cortex charges, and the bill taught them the rules only after they had paid them.

That is the whole trouble with Cortex pricing: the mechanics are unlike anything else in Snowflake, so the numbers feel unpredictable right up until you understand the meter. This guide flips the order. We break down exactly how Snowflake Cortex is priced in 2026, service by service, with real cost examples for common workloads, so you can price an AI feature before you ship it instead of after it startles your finance team.

How Snowflake Cortex Charges You

Cortex breaks the mental model most Snowflake users bring to it. A warehouse bills for how long it runs; Cortex bills for what it processes. Getting three ideas straight up front makes the rest of the pricing obvious.

AI Credits: the 2026 change

Historically Cortex was billed as standard Snowflake credits, priced by edition. On April 1, 2026, Snowflake split AI onto its own meter. AI Credits are now a separate billing currency, flat-priced regardless of your edition: $2.00 per AI Credit if you allow requests to route across regions, and $2.20 if you pin them to your home region for data-residency reasons. Platform Credits for warehouse compute and storage still follow the old $2, $3, and $4 per-credit rates by edition. The practical takeaway is unusual: your AI unit price is set by your data-residency posture, not your Snowflake tier or the model you pick.

Tokens in, tokens out

LLM functions bill per token, and they count input and output separately, with output often running five to eight times the input rate. That asymmetry matters because AI workloads are lopsided: a summarization job might read 5,000 tokens and write 200, while a generation job does the reverse. Which side of that ratio you sit on shapes your bill as much as the volume does. If tokens are new to you, our token economics and TokenOps guide covers the fundamentals.

The two meters running at once

Nearly every Cortex query bills you twice. The AI Credit meter charges for the tokens the model processes. The Platform Credit meter charges for the virtual warehouse that runs the surrounding SQL. Most people watch the first and forget the second, then wonder why a cheap-looking function still moved the warehouse bill. Treat them as two line items from the start.

Cortex Pricing by Service

Cortex is a suite, not a single product, and each service is billed

differently. Here is how each one is billed.

LLM and AISQL functions

Functions like AI_COMPLETE, SUMMARIZE, TRANSLATE, SENTIMENT, and the AISQL family are the workhorses. They bill purely on tokens processed, at effective rates from about $0.12 per million on small, fast models to around $5.10 per million on the largest, most capable ones. For most classification and extraction tasks, a small model is both cheaper and entirely adequate, which is why model selection, not volume, is usually the deciding factor in the bill.

Cortex Search: the meter that never sleeps

Cortex Search powers retrieval and RAG, and it bills on two components. Embedding compute is a per-token charge incurred whenever data is inserted or updated, including the full initial index build. Serving compute is a continuous charge per gigabyte-month of indexed data that runs the entire time the service is resumed, whether or not a single query hits it. That standing fee is the most common source of Cortex surprises, so suspend search services you are not actively using.

Cortex Analyst

Cortex Analyst turns natural language into SQL. Billing depends on how you call it: through the standalone Analyst API it is priced per 1,000 messages, while Snowflake's recommended path, invoking it through Cortex Agents, uses the same token-based AI Credit model as Agents. Either way, running the SQL it generates incurs normal virtual warehouse compute, so the natural-language step is only part of the cost.

Cortex Agents, Document AI, and fine-tuning

Cortex Agents and Snowflake Intelligence orchestrate multi-step workflows and bill in AI Credits per million tokens, with costs adding up across every underlying service the agent invokes, so an agent that calls both Analyst and Search pays for all of it. Document AI carries a high per-document cost that scales quickly on large sets, and fine-tuning consumes credits fast, especially on larger models, so both deserve a tight scope and a resource monitor before you turn them loose.

Real Cost Examples

Numbers make the model concrete. The estimates below are illustrative and depend on your exact model, token counts, and AI Credit rate, but they show the shape of real workloads. They exclude base Snowflake platform costs.

The pattern is clear: lightweight AISQL at scale stays in double digits, generation on large models reaches four figures, and the real danger is not steady usage but a single unbounded call. Before shipping, estimate tokens with the COUNT_TOKENS function on a representative sample, and after shipping, track actual spend through the ACCOUNT_USAGE views such as CORTEX_FUNCTIONS_USAGE_HISTORY and CORTEX_AGENT_USAGE_HISTORY.

What Actually Drives Your Cortex Bill

• Model selection. The largest lever by far. Moving a task from a large model to a small one can cut its cost 10 to 40x, and most routine tasks do not need the flagship.
• Output length. Output tokens cost several times input, so unbounded or verbose generations are a quiet multiplier. Cap them.
• Search serving time. A resumed search service bills per gigabyte-month around the clock. Idle indexes are pure waste.
• The forgotten warehouse. The SQL around a Cortex call still runs on a warehouse. Oversized or always-on warehouses inflate the second meter.
• Re-embedding. Re-indexing unchanged rows repays the embedding cost for nothing. Store embeddings and only re-embed what changed.

How to Keep Cortex Costs Under Control

• Default to the smallest model that clears your quality bar, and require sign-off before anyone switches a workload to a larger one.
• Cap output length on generation functions, and never run an AI function over an unbounded table without a LIMIT or a WHERE clause in testing.
• Suspend Cortex Search services when idle, and only re-embed rows that actually changed.
• Run AI experiments on dedicated warehouses with resource monitors and credit quotas, so a runaway test cannot drain a shared budget.

Because Cortex spend sits inside Snowflake yet behaves like AI spend, it belongs in the same attribution and governance you apply to the rest of your AI and cloud bill. Our Snowflake AI agent cost anomalies and FinOps for AI guides go deeper on attributing and governing exactly this kind of consumption.

The Takeaway

Cortex did not make AI cheaper or more expensive than calling an outside API. What it changed is where the cost lives: inside your database, on two meters, priced by the token and the gigabyte-month rather than by the query. Once that clicks, most of the surprises stop.

Three things are worth carrying with you. Model choice is your biggest single lever, so treat it like warehouse sizing rather than a default you never revisit. The meter people forget, Cortex Search serving, is the one that quietly runs while nobody is looking, so pause what you are not using. And the token is now a unit you budget against, the same way you once budgeted compute hours. Get those three right, and Cortex becomes what it should be: AI you can price to the penny before you build it, not explain away after the invoice lands.