Snowflake Cortex vs Databricks AI Functions: What 1,000 LLM Calls Cost

Summarising 1,000 documents with Claude Sonnet 4.5 costs $9.00 in model tokens on the Anthropic API, and exactly $9.00 on Databricks at its example DBU price. Run the same calls through Snowflake's AI_COMPLETE function in SQL and the reported rate card puts the bill at $10.80, a 20% premium. Snowflake Cortex pricing is not one number. It depends on which door you use to reach the same model.

The short answer on LLM in SQL cost

For a buyer comparing the two warehouses, here is the answer in one paragraph.

Databricks AI Functions bill Claude and other hosted models at rates that convert to the model provider's own list price when you multiply by the DBU price. Snowflake bills AI_COMPLETE in AI Credits at a reported 20% above the direct API for Claude Sonnet 4.5, and its Cortex REST API sits at list price. Both also bill the warehouse that runs your query.

That is the whole finding. The rest of this post shows the working, so you can redo it for your own model, token counts and contract. Rates in this post were last verified on 4 November 2026, and both vendors revise their tables without much notice.

How Snowflake Cortex credits per token convert to dollars

Snowflake moved its AI features onto their own meter during 2026. AI functions such as AI_COMPLETE, AI_CLASSIFY and AI_EMBED are now billed in AI Credits per million tokens, and the rate depends on the model you name in the call. Warehouses, storage and data transfer stay on the older Platform Credit pricing.

AI Credits are a separate meter with a fixed price

The useful change is that AI Credits have a published dollar price. Snowflake's AI pricing documentation lists $2.00 per AI Credit under global routing and $2.20 under regional routing. The price is the same on Standard, Enterprise and Business Critical editions, which removes the edition guesswork that made older Cortex estimates unreliable.

Two details on that page matter for your budget. Automatic discounts apply based on annual contract value, so a large account may pay below these figures. And capacity discounts on Platform Credits do not carry over to AI Credits, so a negotiated warehouse discount tells you nothing about your AI rate.

Regional routing is the setting a regulated team usually picks, because it keeps inference inside a named cloud region. That choice costs a flat 10% on every AI Credit. If your data residency policy forces it, put the $2.20 figure into the model from the start.

Three surfaces, three rates for one model

Snowflake sells access to the same Claude model through more than one door, and the doors are priced differently. Snowflake's own Cortex REST API billing guide, updated in April 2026, lists claude-sonnet-4-5 at $3.00 per million input tokens and $15.00 per million output tokens. That matches Anthropic's list price exactly.

The SQL function is dearer. An October 2026 analysis that quotes the Service Consumption Table directly gives AI_COMPLETE a rate of 1.80 AI Credits per million input tokens and 9.00 per million output tokens for the same model. At $2.00 per credit that is $3.60 and $18.00. Cortex Agents and Snowflake Intelligence sit higher again in the same source, at 1.95 and 9.76 credits.

So the cheapest way to reach Claude inside Snowflake is not the SQL function the product is marketed around. It is the REST endpoint your application code would call anyway. That is worth knowing before you write a pipeline around AI_COMPLETE because it was the first example in the documentation.

The AI_COMPLETE row rests on a secondary reading of Snowflake's consumption table and should be checked against the current PDF before you sign anything. Every other row comes from the vendor's own published page. The Databricks row assumes the $0.07 example price shown on the pricing page for AWS US East.

How Databricks DBU AI pricing converts to dollars

Databricks bills model inference in Databricks Units, its general compute currency. For hosted foundation models on the pay-per-token mode, the pricing page lists a DBU rate per million input tokens and per million output tokens, then tells you to multiply by your SKU price per DBU.

The DBU rates are list prices in disguise

Run the multiplication and the odd-looking decimals resolve into round numbers. The proprietary foundation model pricing page lists Claude Sonnet 4.5 at 42.857 DBU per million input tokens and 214.286 per million output tokens. At the page's example price of $0.07 per DBU, that is $3.00 and $15.00, identical to Anthropic.

The pattern holds across the catalogue. Claude Haiku 4.5 at 14.286 and 71.429 DBU converts to $1.00 and $5.00. GPT-5 at 17.857 and 142.857 DBU converts to $1.25 and $10.00. Databricks shows its own worked example for Kimi K3, where 42.857 DBU at $0.07 becomes $3.00 per million input tokens, so the conversion is the vendor's method, not ours.

The design choice is clear. Databricks treats the model as a pass-through and earns on the compute around it. If your DBU price differs from $0.07, every model line moves by the same ratio, so the relative comparison in this post still holds.

There is one surcharge to watch. The page states that when regional processing for data residency is enabled, models carrying its marker take a 10% uplift. That is the same size as Snowflake's regional premium, so a residency requirement does not by itself favour either vendor.

ai_summarize hides the model choice

The convenience functions are a different matter. Databricks' AI Functions documentation, updated in October 2026, says task functions such as ai_summarize and ai_classify run inference on Databricks-managed serverless infrastructure and are billed in addition to your query compute. The ai_summarize reference page names no model. It says only that the underlying models are Apache 2.0 licensed and that Databricks might change the model.

You cannot convert that to a per-token dollar figure from the public rate card, because you do not know which row of the rate card applies. Usage lands in the billing system table under the Model Serving product with a batch inference offering type, so you can measure it after the fact. You cannot price it before you run it.

My recommendation is to use ai_query with a named endpoint for anything you need to budget. It costs the same per token as the rate card says, and it pins the model, which also protects your output quality from a silent swap.

The conversion table: dollars per 1,000 summarisation calls

Per-million-token rates are hard to reason about. A per-call figure is easier, so this section fixes a workload and prices it everywhere.

The workload is a plain summarisation job. Each call sends 2,000 input tokens, roughly a long support ticket or a short contract clause with its instructions. Each call returns 200 output tokens, a summary of a few sentences. So 1,000 calls consume 2 million input tokens and 0.2 million output tokens.

Computed by Zan Digital from the published rates in the sources listed below, retrieved 4 November 2026. Regional figures apply Snowflake's $2.20 AI Credit price. Haiku 4.5 through AI_COMPLETE is left out because no source we could open gave its AI Credit rate.

Three things stand out. Databricks and the direct API are tied to the cent. Snowflake's REST path is tied too, until regional routing adds 10%. And the cheapest option on the whole chart is not a warehouse at all, it is Anthropic's Batch API at half price, which suits overnight summarisation perfectly.

If you already compare hosted model prices across providers, the same logic applies here. The inference provider rate card comparison found identical list prices hiding very different bills, and the warehouse case is a variant of that pattern. The discounts that move the total most are covered in the breakdown of prompt caching and batch discount rates.

Where the warehouse markup actually comes from

The warehouse markup has two sources. One is on the rate card and one is not, and the second one is the one most teams forget.

The SQL convenience premium

The first source is the per-token rate itself. Snowflake prices AI_COMPLETE 20% above its own REST endpoint for Sonnet 4.5, on the second-hand reading of Table 6 used here. On 1,000 calls that is $1.80. On 1 million calls a month it is $1,800 a month for the privilege of writing the call in SQL.

I think that premium is defensible for exploration and indefensible for a production batch job. Exploration is where a SQL function saves real analyst time. A nightly job that summarises the same table every day is engineering work, and it should take the cheaper door.

The warehouse that stays awake

The second source is not on the AI rate card at all. Snowflake's cost guidance for AI functions says the cost of keeping a warehouse active continues to apply while a query calls an AI function. It also recommends a warehouse no larger than Medium for these calls, because a bigger warehouse does not make the model respond faster.

A Medium standard warehouse burns 4 Platform Credits per hour, billed per second with a 60-second minimum each time it starts. The model calls are slow relative to ordinary SQL, so the warehouse sits mostly idle waiting on inference. On a Medium warehouse, every 15 minutes the job runs costs 1 Platform Credit on top of the AI Credits.

Compare that with the token bill. The whole 1,000-call Sonnet job above costs 5.4 AI Credits. If a slow model run holds a Medium warehouse open for an hour, it adds 4 Platform Credits, priced at whatever your edition contract says. Warehouse time can therefore be a large share of the total, and it never shows up in a per-token comparison.

Databricks has the same structure. Its AI Functions page states that the compute running your query, such as a SQL warehouse or job cluster, is always billed, and the model inference is billed on top. Neither vendor's per-token figure is the full cost of an LLM call made from SQL.

Open-weight models flip the Snowflake vs Databricks comparison

On Claude, Databricks is never more expensive per token than Snowflake. On open-weight models the answer depends on the shape of your workload, because the two vendors price input and output differently.

Databricks lists Llama 3.3 70B at 7.143 DBU in and 21.429 DBU out per million tokens, which is $0.50 and $1.50 at $0.07. Snowflake's REST billing guide lists llama3.3-70b at a flat $0.72 for both input and output. One vendor charges three times more for output than input. The other charges the same for both.

The break-even is simple algebra. With 2,000 input tokens per call, the two lines cross at about 564 output tokens, which is an output-to-input ratio of roughly 28%. Below that, Databricks is cheaper. Above it, Snowflake is.

Summarisation, classification and extraction all live well below that ratio. Our 200-token summary job costs $1.30 on Databricks and $1.58 on Snowflake, so Databricks wins by about 18%. Writing product descriptions from a short spec sheet sits above it, and Snowflake wins. Pick the model rate card that matches what your job actually does, not the one with the lowest input price. The same input-versus-output trap shows up in the per-task cost comparison of small language models.

Where this comparison is weakest

Three parts of this analysis rest on less than a primary document I could open and read in full. Here they are, with what would change the conclusion.

The AI_COMPLETE rate is second-hand, and so is the DBU price

The 1.80 and 9.00 AI Credit figures for AI_COMPLETE come from an October 2026 article that quotes Snowflake's Service Consumption Table. The table itself is a PDF that our tools could not extract. If the current table shows a lower rate, the 20% premium shrinks. Search results also surfaced an older January 2026 version of the table with lower figures, which suggests the rate has moved during the year, so check the effective date on whatever copy you read.

The Databricks side has a similar soft spot. The pricing page uses $0.07 per DBU as an example for AWS US East, and the live SKU tiles did not load for us. Your negotiated DBU price, cloud and region decide the real number. Because every model row scales by the same ratio, a different DBU price changes the dollar totals but not which model is cheaper. It does change the Databricks-versus-Snowflake gap, and at a DBU price much above $0.07 Databricks loses its tie with list price.

Where the warehouse genuinely wins

The strongest argument against this post is that per-token cost is the wrong comparison. Calling a model from SQL keeps the data inside the warehouse's governance boundary. Row-level access policies, masking and audit logs apply to the query that feeds the model. Exporting the same rows to an application that calls the Anthropic API means rebuilding those controls, or quietly skipping them.

For a regulated team, 20% on a $9.00 job is cheap insurance against a data handling incident. I would still move high-volume production jobs to the REST or batch path, but I would not argue with a compliance lead who prefers to keep everything in SQL. The governance case is made in more depth in the comparison of governance controls in text-to-SQL tools, and it applies here too.

One more caveat. The workload of 2,000 input and 200 output tokens is a choice, not a measurement. Your own token counts will differ, and the Llama crossover shows that the shape of the workload can reverse the ranking. Treat every per-call figure here as a worked example of the method.

Frequently asked questions

Where to start this week

Pull one real number before you change anything. In Snowflake, query the AI function usage views for the last 30 days and split the credits by model and by function name. In Databricks, filter the billing usage table to the Model Serving product. You want the share of your AI spend that runs through the SQL convenience path.

Then take your single largest recurring job and reprice it three ways using the tables above: the SQL function, the REST or ai_query path with a named model, and the provider's batch API. If the gap is more than a few hundred dollars a month, move that one job and leave the exploratory queries where they are. For a wider view of why AI bills drift away from the rate card, read the guide to why AI credits and tokens confuse buyers.

Related reading