Snowflake Cortex Analyst Pricing, Snowflake Cortex Pricing and the Real Cost per Question

Snowflake Cortex Analyst costs 67 Platform Credits per 1,000 messages when you call it through its API, according to Snowflake's Service Consumption Table effective 30 September 2026. At on-demand prices in AWS US East that is $0.134 a question on Standard, $0.201 on Enterprise, $0.268 on Business Critical and $0.402 on VPS. Only successful responses are billed. The warehouse that then runs the generated SQL is billed separately, and if you reach Cortex Analyst through Cortex Agents, which Snowflake now recommends, you pay for model tokens in AI Credits instead of per message.

That makes Cortex Analyst one of the few AI features in this market with a published unit price, which is useful. It also means the bill has three moving parts, and most cost estimates online quote only one of them. Below is each part, the monthly math at three team sizes, and how the price per question compares with the alternatives a Snowflake customer is likely to look at.

How much does Snowflake Cortex Analyst cost?

The message price is fixed in credits, so the dollar figure depends on your edition and region. These are the on-demand Platform Credit prices for AWS US East (Northern Virginia) from Table 2(a) of the consumption table, multiplied by 0.067 credits per message.

Two details change the number. First, regions outside the cheapest US ones cost more per credit: the same Enterprise question is $0.261 in EU Dublin at $3.90 a credit. Second, because the API is billed in Platform Credits, a capacity contract discount does apply to it, so a customer with 20 percent off warehouse credits also gets 20 percent off Cortex Analyst API messages. That is not true of the other route, which matters more than it looks. Our Snowflake pricing guide covers editions, storage and capacity contracts in full.

Is Cortex Analyst billed per message or per token?

Per message through the API, per token through agents. Footnote 23 of the consumption table says the per-message price "will only be applicable when using the Cortex Analyst API." When Cortex Analyst runs as a tool inside Cortex Agents or Snowflake CoWork, Table 6(d) applies: the model's input and output tokens are billed in AI Credits per million tokens, at a rate that depends on the model the agent uses.

AI Credits are a separate currency. On demand they cost $2.00 if you allow cross-region inference and $2.20 if you pin requests to your home region, and the price is the same on every edition. Snowflake's table states that "the Platform Credit Discount does not apply to AI Credits," so your warehouse discount does not reach them. Large contracts get their own small AI Credit discount instead, from $1.96 above $1.2 million a year down to $1.88 above $40 million.

What a question costs on the agent route depends on how many tokens it uses, which depends on the size of your semantic model and how many steps the agent takes. As an illustration only, a question that sends 15,000 input tokens and gets 1,500 output tokens back costs about 0.044 AI Credits on claude-sonnet-4-5 (1.95 and 9.76 AI Credits per million) and about 0.015 on claude-haiku-4-5 (0.65 and 3.25). At $2.00 a credit that is roughly $0.09 and $0.03 a question. Measure your own token counts before relying on either figure.

The practical reading: on Business Critical or VPS, where a Platform Credit costs $4.00 to $6.00, the agent route with a small model is likely to be cheaper per question. On Standard with a negotiated discount, the API's fixed message price can be the cheaper and more predictable one. Snowflake's documentation recommends moving to Cortex Agents for answer quality, so check which route your team is actually calling before you budget.

Does Cortex Analyst use warehouse credits?

Yes, and this is the part most estimates leave out. Snowflake's Cortex Analyst documentation says the charges "cover AI costs for text-to-SQL" and that "additional warehouse costs apply when you execute the SQL generated by Cortex Analyst." The generated query runs on a virtual warehouse like any other query.

The consumption table sets the rules for that warehouse. An XS standard warehouse uses 1 Platform Credit an hour, each size up doubles it, and every time a suspended warehouse resumes it bills a minimum of one minute, then per second. For plain-English questions that arrive one at a time through the day, the one-minute minimum dominates: a two-second query on a warehouse that had suspended still costs a full minute. On Enterprise at $3.00 a credit, an XS warehouse that stays awake four hours each working day to answer questions is 84 credits a month, or $252.

What does Cortex Analyst cost per month?

Here is the same company at three sizes, assuming five questions per person per working day, 21 working days, Enterprise edition, AWS US East, on-demand list prices and the API route. The warehouse line uses the four-hours-a-day XS assumption above and will vary with your workload.

Three costs sit outside this table. Cortex Analyst answers from a semantic model or semantic view that someone has to write and maintain, and that is analyst time, not credits. If you add Cortex Search so it can match literal values such as product names, that bills 6.3 AI Credits per GB of indexed data a month. And Cortex Analyst runs natively in a limited set of regions, nine in Snowflake's documentation, with cross-region inference for the rest. Because AI Credits are a new line on the invoice, it pays to track AI credits next to the rest of your cloud spend from the first month rather than discovering them at renewal.

How Cortex Analyst pricing compares with other ways to ask questions in plain English

Every serious analytics vendor now sells some version of "type a question, get an answer," and they bill it in completely different units. Normalized to a cost per question at list price, the spread is large.

The ThoughtSpot figure divides the Pro seat price by its 25-query allowance, which is how ThoughtSpot publishes it; our ThoughtSpot vs Tableau comparison works through that tier at four team sizes. The Copilot route is covered in our Power BI Copilot license guide, and Databricks' equivalent, Genie, has its own DBU-based meter that we break down in Databricks Genie pricing. If you are still choosing the warehouse itself, Snowflake vs Databricks prices Cortex Analyst against Genie side by side.

Is there a cheaper alternative to Snowflake Cortex Analyst?

For one specific job, yes: business users asking questions about tables in Snowflake and getting a number back with the SQL that produced it. We built AgentSQL for that job, so weigh this section accordingly. It connects to Snowflake with a read-only role, reads the live schema, writes Snowflake SQL, runs it, and returns the answer, a chart and the query. The Team plan includes Snowflake and BigQuery connections, up to 10 seats and 1,500 questions a month for $59 billed yearly, which works out to about 4 cents a question, roughly a fifth of the Enterprise API price.

Two honest caveats. Your warehouse still runs every query, so the warehouse half of the bill is the same whichever tool writes the SQL; size the warehouse and its auto-suspend the same way. And Cortex Analyst has advantages a third-party tool cannot match: it stays inside Snowflake's governance boundary, and a well-built semantic model encodes business definitions that a schema alone does not. If your metrics need that layer, Cortex Analyst earns its price. If your questions are mostly counts, sums and trends over clean tables, a flat-priced tool is the cheaper way to answer them. See how the connection works on AgentSQL for Snowflake.

How to keep Cortex Analyst costs predictable

1. Find out which route you are on. API calls show up in CORTEX_ANALYST_USAGE_HISTORY; agent calls bill tokens. The same question can cost several times more or less depending on the answer.
2. Put the generated SQL on its own warehouse. A dedicated XS warehouse with a short auto-suspend and a resource monitor separates question traffic from ETL and makes the warehouse half of the bill visible.
3. Watch METERING_HISTORY for AI_SERVICES weekly. AI Credits sit outside your Platform Credit discount, so a pilot that grows quietly shows up at list price.
4. Price the semantic model work. Count the tables and metrics the model must cover and the hours to keep it current. That cost is real even though it never appears on the Snowflake invoice.
5. Re-read the consumption table each quarter. Snowflake revised it twice in September 2026 alone, and the Cortex rows are where changes land.

The quickest way to see whether your team's questions need a semantic layer at all is to ask a few of them. Type one into the console at the top of this page and read the SQL it writes; if the answers are right without a model, you have your price comparison.