Snowflake Cortex AI Functions pricing: AI Credit rates per million tokens
Snowflake bills Cortex AI Functions such as AI_CLASSIFY, AI_SENTIMENT, AI_EXTRACT and AI_COMPLETE in AI Credits per million tokens, and each AI Credit costs $2.00 with global routing or $2.20 with regional routing. Those two numbers come from Snowflake's AI pricing documentation and from the Snowflake Service Consumption Table effective November 12, 2026, both checked on November 14, 2026. The per-function rates sit in Table 6(a) of that consumption table: AI_CLASSIFY and AI_FILTER cost 1.62 AI Credits per million tokens, AI_SENTIMENT 1.60, AI_TRANSLATE 1.63, AI_EXTRACT 5.55, and AI_COMPLETE depends on the model you pick.
The rates are most relevant to data teams that want to run classification, sentiment scoring, extraction or summarization directly in SQL over warehouse tables, because the charge scales with every row and every token. A query that classifies 200,000 support tickets is billed for the ticket text, for the category labels repeated on each row and for a hidden prompt that Snowflake adds to the input.
AI Credits are a flat price that does not change with edition (Standard, Enterprise, Business Critical or VPS) or with region, according to Snowflake's AI pricing page. The only lever on the price of one AI Credit is the routing setting, plus automatic discounts tied to annual contract value. In the illustrative month worked out below, 200,000 tickets classified and summarized with a mid-priced model come to about $348 with global routing.
··········
HOW SNOWFLAKE BILLS CORTEX AI FUNCTIONS.
Snowflake's cost documentation states that Cortex AI Functions incur compute cost based on the number of tokens processed, and that roughly four characters of English text make one token, with the exact ratio varying by model. The billing basis differs by function, and the documentation splits them into groups.
Text-generating functions, which the documentation lists as AI_COMPLETE, AI_CLASSIFY, AI_FILTER, AI_AGG, AI_SUMMARIZE and AI_TRANSLATE, are billed on input and output tokens. Embedding functions such as AI_EMBED and AI_SIMILARITY are billed on input tokens only. AI_EXTRACT and AI_REDACT are billed on input and output tokens, and for AI_EXTRACT each document page counts as 970 input tokens. AI_PARSE_DOCUMENT is billed per page processed instead of per token, and audio is counted at 50 tokens per second.
Several task functions add their own prompt to your text before it reaches the model. Snowflake lists AI_CLASSIFY, AI_FILTER, AI_AGG, AI_SENTIMENT, AI_SUMMARIZE_AGG and AI_TRANSLATE among them, so the billed token count is always higher than the length of the column you pass in. For AI_CLASSIFY the documentation adds that labels, label descriptions and examples are billed as input tokens for each record processed, not once per call.
Warehouse time is a second, separate charge. A query that calls a Cortex function keeps its warehouse running, and that warehouse consumes Platform Credits at its normal hourly rate. Snowflake recommends a warehouse no larger than MEDIUM for Cortex AI Functions, because a larger warehouse does not improve their performance and can add cost.
··········
CORTEX AI FUNCTIONS CREDIT RATES IN TABLE 6(A).
The table below lists the rates printed in Table 6(a) of the Snowflake Service Consumption Table effective November 12, 2026, as read on November 14, 2026. Task functions carry a single rate per million tokens; whether that rate applies to input only or to input plus output is set by the documentation described above.
........
Cortex AI Function | AI Credits per 1M tokens | Billed on |
AI_CLASSIFY | 1.62 | Input and output |
AI_FILTER | 1.62 | Input and output |
AI_SENTIMENT | 1.60 | Tokens incl. added prompt |
AI_TRANSLATE | 1.63 | Input and output |
AI_AGG | 1.85 | Input and output |
AI_SUMMARIZE_AGG | 1.85 | Tokens incl. added prompt |
AI_EXTRACT | 5.55 | Input and output |
AI_REDACT | 0.69 | Input and output |
AI_EMBED, snowflake-arctic-embed-m-v1.5 | 0.03 | Input only |
AI_EMBED, snowflake-arctic-embed-l-v2.0 | 0.05 | Input only |
........
AI_PARSE_DOCUMENT is priced per 1,000 pages in Table 6(g): 0.68 AI Credits in OCR mode, 3.66 in Layout mode and 8.18 in AGENTIC mode. Document-heavy pipelines that parse contracts or invoices before extracting fields therefore pay twice, once per page for parsing and once per token for the extraction step.
AI_COMPLETE has separate input, output and cached-input rates for each model. A handful of commonly used models shows how wide the spread is.
........
AI_COMPLETE model | Input | Output | Cached input | Context window |
llama3.1-8b | 0.132 | 0.132 | Not listed | 128,000 |
llama3.3-70b | 0.432 | 0.432 | Not listed | 128,000 |
claude-haiku-4-5 | 0.60 | 3.00 | 0.060 | 200,000 |
mistral-large2 | 1.20 | 3.60 | Not listed | 128,000 |
openai-gpt-4.1 | 1.20 | 4.80 | 0.300 | 128,000 |
claude-sonnet-4-5 | 1.80 | 9.00 | 0.180 | 200,000 |
........
Rates are in AI Credits per million tokens; context windows come from Snowflake's models and regional availability page. The consumption table also carries footnotes stating that some newer models are on promotional pricing that will rise on dates listed there, including November 21, 2026, December 1, 2026 and January 1, 2027.
··········
AI CREDIT PRICE: $2.00 GLOBAL VS $2.20 REGIONAL ROUTING.
The dollar price of an AI Credit depends on the CORTEX_ENABLED_CROSS_REGION account parameter. Snowflake's AI pricing page lists $2.00 per AI Credit for global routing, which applies when the parameter is ANY_REGION, AWS_GLOBAL, GCP_GLOBAL or AZURE_GLOBAL. Regional routing costs $2.20 per AI Credit and applies when the parameter is DISABLED or set to a geography such as AWS_US, AWS_EU, AZURE_US or AZURE_EU, which keeps processing in the home region or in the selected geography.
Only the ACCOUNTADMIN role can change the parameter, and only at account level, with a statement such as the one below taken from Snowflake's cross-region inference documentation.
ALTER ACCOUNT SET CORTEX_ENABLED_CROSS_REGION = 'ANY_REGION';According to that documentation, new organizations created after March 9, 2026 default to ANY_REGION, while other commercial accounts inherit a same-cloud geography value such as AWS_US or AZURE_EU. Customer data stays stored in the home region, prompts and responses travel to the processing region only for the duration of processing, and Snowflake charges no data egress fee for cross-region inference. The choice is a trade between a 10% higher AI Credit price and tighter control over where inference runs, which is often a data residency or compliance requirement.
Table 2(b) of the consumption table also lists lower per-credit prices at higher annual contract tiers, down to $1.88 global and $2.07 regional in the highest band shown. Snowflake states that these ACV-based discounts apply automatically and that capacity discounts do not apply to AI Credits.
··········
WORKED CORTEX COST EXAMPLE FOR A SUPPORT ANALYTICS TEAM.
The figures below are illustrative. A support analytics team wants to classify 200,000 new tickets a month into four categories with AI_CLASSIFY and write a one-sentence summary of each ticket with AI_COMPLETE. After sampling with AI_COUNT_TOKENS, the team measures an average of 320 billed tokens per ticket for classification, including labels and Snowflake's added prompt, and 250 input tokens plus 60 output tokens per ticket for the summary. The queries run on an X-Small warehouse for three hours in total; Snowflake documents an X-Small Gen1 warehouse at 1 credit per hour, and the Platform Credit price of $3.00 is the illustrative Enterprise figure used on Snowflake's own pricing page.
........
Cost line (illustrative) | Volume | AI or Platform Credits | Global routing | Regional routing |
AI_CLASSIFY at 1.62 | 64M tokens | 103.68 AI Credits | $207.36 | $228.10 |
AI_COMPLETE claude-haiku-4-5 | 50M in, 12M out | 66.00 AI Credits | $132.00 | $145.20 |
X-Small warehouse | 3 hours | 3 Platform Credits | $9.00 | $9.00 |
Monthly total | 200,000 tickets | 169.68 AI + 3 Platform | $348.36 | $382.30 |
........
The model choice for the summary step moves the bill more than anything else. The same 50 million input tokens and 12 million output tokens cost very different amounts depending on the model passed to AI_COMPLETE.
........
Summary model (illustrative) | AI Credits | Global routing | Regional routing |
llama3.1-8b | 8.18 | $16.37 | $18.00 |
claude-haiku-4-5 | 66.00 | $132.00 | $145.20 |
openai-gpt-4.1 | 117.60 | $235.20 | $258.72 |
claude-sonnet-4-5 | 198.00 | $396.00 | $435.60 |
........
Output tokens weigh heavily on the premium models, where the output rate is four to five times the input rate, so asking for a 20-word summary instead of a paragraph is a direct cost control. Readers who want the Platform Credit side in more detail can compare it with our breakdown of Snowflake pricing by credits, editions, warehouse sizes and storage.
··········
ESTIMATING AND TRACKING CORTEX AI CREDITS IN SQL.
AI_COUNT_TOKENS returns the estimated number of input tokens for a given function and input, and Snowflake bills it only for compute, not by token count. The query below follows the documented syntax for AI_CLASSIFY, where the categories are passed in the same form as in the real call.
SELECT AVG(AI_COUNT_TOKENS('ai_classify', ticket_text, [
{'label': 'billing'},
{'label': 'outage'},
{'label': 'feature request'},
{'label': 'other'}
])) AS avg_input_tokens
FROM support_tickets SAMPLE (1000 ROWS);The estimate covers input tokens only. Snowflake notes that output tokens are billed but not counted, that AI_TRANSLATE, AI_SUMMARIZE and AI_REDACT estimates cover only the source text, and that structured output on Claude models adds billed input that the estimate does not include. Accuracy varies by model, with published mean absolute percentage errors from about 2% to about 14%.
After the workload runs, the CORTEX_AI_FUNCTIONS_USAGE_HISTORY view in ACCOUNT_USAGE records credits per function, model, query, user and role in one-hour windows, with token or page counts in its METRICS column. Its CREDITS column excludes warehouse credits. The older CORTEX_FUNCTIONS_QUERY_USAGE_HISTORY view is no longer updated, according to Snowflake's documentation.
SELECT function_name,
model_name,
SUM(credits) AS ai_credits
FROM SNOWFLAKE.ACCOUNT_USAGE.CORTEX_AI_FUNCTIONS_USAGE_HISTORY
WHERE start_time >= DATEADD(day, -30, CURRENT_TIMESTAMP())
GROUP BY function_name, model_name
ORDER BY ai_credits DESC;Multiplying the AI Credits from that query by $2.00 or $2.20, depending on the routing setting, gives the list-price spend before any contract discount. The same token logic applies to AI features in other warehouses, as our look at Gemini in BigQuery pricing per million tokens shows for Google's side.
··········
COMMON MISREADINGS OF CORTEX AI PRICING.
The first misreading is to treat AI Credits like Platform Credits. Snowflake's pricing page states that AI Credit pricing applies only to the listed AI features, while warehouses, storage and data transfer stay on Platform Credit pricing, whose price varies by edition and region.
The second is to budget AI_CLASSIFY from the length of the text column alone. Every label, description and example is billed again on each row, and Snowflake's own AI_COUNT_TOKENS examples show a short sentence growing from 187 tokens with four bare labels to 298 tokens once descriptions, a task description and one example are added.
The third is to assume that a function available in a region makes every model available there. Snowflake's regional availability page warns that a function being available in your region does not mean every model it accepts is, and many models are reachable only through cross-region inference. An account set to DISABLED may not reach the model it planned to use; widening the setting to a geography such as AWS_US keeps the regional $2.20 price, while a global setting moves it to $2.00. Document workloads deserve a separate check, as our report on Snowflake AI_EXTRACT handling 1,000-page documents describes for the extraction function.
··········
SNOWFLAKE CORTEX AI FUNCTIONS COSTS: WHAT TO CONFIRM IN YOUR ACCOUNT.
Snowflake publishes the full set of facts needed to price a Cortex workload: per-function and per-model rates in AI Credits per million tokens in Table 6(a), page rates for AI_PARSE_DOCUMENT in Table 6(g), and the $2.00 and $2.20 AI Credit prices tied to the routing parameter. The token counts are the uncertain part, and AI_COUNT_TOKENS on a sample, followed by the CORTEX_AI_FUNCTIONS_USAGE_HISTORY view after a pilot run, replaces assumptions with measured figures.
Before a large backfill, an administrator should check the current value of CORTEX_ENABLED_CROSS_REGION, confirm that the chosen model is reachable under that setting, read any promotional-pricing footnote on the model, and keep the warehouse at MEDIUM or smaller. Contract discounts on AI Credits depend on annual contract value and are not shown in the public list prices, so the final dollar figure for a given company has to come from its own Snowflake agreement.
··········
FOLLOW US FOR MORE.
·····
DATA STUDIOS
·····
datastudios.org
