Snowflake changed how it bills for AI on April 1, 2026. If your team is running Cortex functions, Cortex Search, Cortex Agents, or Snowflake Intelligence, the pricing model you learned last year is already out of date. Here's what Cortex costs today, how the math actually works, and how it stacks up against Databricks.
How Snowflake Cortex pricing works now
Snowflake introduced AI Credits as a separate billing currency, decoupled from your Snowflake edition. AI Credits are flat priced:
- $2.00 per credit if you allow global routing (cross-region inference enabled)
- $2.20 per credit if you pin requests to your home region for data residency
This applies to Cortex Agents, Cortex Code, and Snowflake Intelligence. Cortex AI Functions, the SQL-callable layer like AI_COMPLETE and AI_EMBED, still bill in standard edition-priced Snowflake Credits, not the new AI Credit currency. That means the same function can cost a different amount depending on which surface you call it through.
On top of AI Credits, three more charges show up on a Cortex bill:
- Cortex Search serving, billed per GB of indexed data per month, and it runs continuously whether or not anyone queries it
- Document AI, billed per document processed
- Standard warehouse compute, since any SQL that runs alongside a Cortex call still burns regular credits
Per-model rates for AI Functions range roughly from $0.12 per million tokens on small open-source models up to $5.10 per million tokens on the most capable frontier models. Model choice is the single biggest lever on your bill. The exact per-model, per-function rate lives in Snowflake's Service Consumption Table, not the headline pricing page, so budgeting by memory or by an old quote is a good way to get surprised.
Real cost examples
Example 1: A single AI_COMPLETE batch job One team ran AI_COMPLETE across roughly a million rows, using the most capable frontier model available, with a 1,000-token prompt and about 3,000 tokens of output per row. Total cost: north of 26,000 AI Credits, or roughly $52,000 at the $2.00 global rate. The identical workload on a smaller open-source model came in around 440 credits, under $900. Same rows, same prompt, a 59x difference in price for the model swap alone.
Example 2: A moderate monthly Cortex Agents workload A team running Cortex Agents on Claude Sonnet 4.6, with a mix of generation, cached input, bulk classification, code assistance, and analyst queries, saw a bill like this after the April 2026 changes took effect:
| Workload | Volume | Monthly cost |
|---|---|---|
| Cortex Agents generation | 50M output tokens | ~$976 |
| Cortex Agents input, 60% cache hit rate | 200M input tokens | ~$546 |
| Bulk classification, small open-source model | 500M tokens | ~$286 |
| Cortex Code | 10M output tokens | ~$165 |
| Cortex Analyst, routed standalone | 5K queries, ~2M tokens | ~$39 |
| Total | ~$2,012/month |
The same workload before the April pricing changes ran closer to $6,630/month. The gap comes from AI Credits decoupling from edition pricing, prompt caching cutting repeated-context cost, and routing bulk work to cheaper models instead of defaulting to the frontier model for everything.
Example 3: Scaling up to a heavy production workload Take that same mix and scale it 10x for a company running Cortex across multiple product lines: 500M output tokens, 2B cached input tokens, 5B bulk classification tokens, 100M code tokens, and 50K analyst queries a month. Directionally, that lands in the $18,000 to $22,000 a month range post-optimization, and would have run $60,000+ under the pre-April model. At that scale, warehouse compute and Cortex Search serving costs start to matter as much as the AI Credit line itself, and they don't show up in most people's mental model of "the AI bill."
The pattern holds at every scale: model selection, caching, and routing are worth 60 to 70% of your Cortex bill. Most teams don't have visibility into which of their workloads are eligible for those savings until they can see cost broken down by model, by function, and by team.
Snowflake Cortex vs. Databricks: how the pricing models compare
Databricks doesn't bill AI the way Snowflake does. Instead of a token-priced AI Credit, Databricks runs everything, including Mosaic AI Foundation Model Serving, through its Databricks Unit (DBU) system.
- CPU-based model serving starts around $0.07 to $0.08 per DBU
- GPU-based serving varies heavily by GPU class, and a single large endpoint running continuously can generate tens of thousands of dollars a month on its own
- Foundation Model APIs offer pay-per-token pricing for open-source models, similar in spirit to Snowflake's AI Credit model
- Databricks also layers a separate cloud infrastructure bill on top of the DBU charge. All-in cost typically runs 50 to 100% higher than the DBU line item alone once you add the underlying AWS, Azure, or GCP compute
The practical difference: Snowflake's AI Credit pricing is simpler to forecast per token, but stacks a second warehouse-compute meter on top for anything beyond the raw model call. Databricks is more transparent about compute-per-second, but GPU-backed serving costs are far less predictable and idle endpoints keep burning DBUs unless you configure auto-suspend. Neither vendor makes it easy to see the true blended cost of a workload without pulling usage data out of native consumption views and reconciling it against actual line items.
For teams running both platforms, or comparing the two before a big commitment, the real question isn't which sticker price is lower. It's which platform gives you accurate, real-time visibility into what a specific model, team, or product feature is actually costing you, in dollars, not credits or DBUs.
Where Finout fits
Snowflake Cortex and Databricks both bill in their own proprietary units, credits on one side, DBUs on the other, and neither one gives you a native way to see that spend next to your Anthropic API bill, your OpenAI bill, or the rest of your cloud and SaaS stack.
Finout sits on top of all of it. Built on Finout's patented MegaBill data layer, Finout ingests and normalizes Cortex AI Credit consumption, Databricks DBU consumption, and direct model provider spend from Anthropic, OpenAI, and others, alongside your core AWS, Azure, GCP, and Kubernetes costs, all in one place. That means:
- Cortex spend broken down by model, function, team, and workload, not just a lump AI Credit line
- Cost per Cortex Agent, per Cortex Search index, or per Databricks endpoint, mapped to the product feature or team that's driving it
- One view across every AI vendor you use, so a Claude Sonnet call through Cortex Agents and a direct Anthropic API call show up in the same unit economics, not two disconnected bills
- Budgets and anomaly detection that catch a 26,000-credit AI_COMPLETE job before it lands on next month's invoice, not after
Finout scales the same way whether you're running Cortex on one warehouse or across a dozen product lines and multiple clouds. You get one system of record for AI, cloud, and SaaS cost, instead of stitching together Snowflake's consumption views, Databricks' usage dashboards, and provider invoices by hand every month end.
cloud & AI spend
.svg)
