Snowflake Cortex vs Databricks Mosaic AI vs BigQuery AI: Real Enterprise AI Cost per 1 Billion Tokens in 2026

The short answer: all three "AI in your data warehouse" platforms bill through the warehouse's own currency, and that currency's real dollar value changed materially for Snowflake in 2026. Snowflake's Cortex AI features moved to a flat, edition-independent AI Credit priced at $2.00 (global routing) or $2.20 (regional) on April 1, 2026 — a genuine repricing that materially reduced effective AI costs for customers whose previous edition-priced credits were substantially above the new flat AI Credit rate, since AI Credits no longer inherit a Business Critical or VPS edition's premium per-credit rate. Databricks bills AI through its existing DBU (Databricks Unit) currency at a reported $0.07 per DBU for foundation model serving, and its own published per-model DBU rates work out to prices that, for at least one open-weight model tested, are cheaper than a dedicated inference provider's direct rate. BigQuery's AI functions don't have a standalone per-token rate at all — they bill BigQuery compute for data processed plus a separate Vertex AI charge for the model call itself, at Vertex's batch-API rate for any Gemini 2.0-or-later model.

What changed: Snowflake's April 2026 AI Credit repricing

Before April 1, 2026, Cortex AI functions billed in ordinary Snowflake Credits, whose price varied by edition and region — from $2.00 on Standard AWS US East up to $9.30 on VPS Switzerland, or as high as $5.20 for Business Critical in Azure West Europe. Snowflake introduced AI Credits as a separate, flat-priced currency decoupled from edition: $2.00 per credit globally, $2.20 if pinned to a home region for data residency. Snowflake's own current documentation confirms this AI Credit pricing applies broadly across its AI surface — Cortex Agents, Cortex Code, Snowflake Intelligence, the Cortex REST API, AI Functions (AI_COMPLETE, AI_EMBED, AI_CLASSIFY, AI_EXTRACT), Cortex Search, Cortex Batch Search, and AI Parse Doc all bill in AI Credits, not the older edition-priced Platform Credits. Two features remain on the legacy Platform Credit model: Cortex Fine-tuning and the standalone Cortex Analyst API (Cortex Analyst invoked through Cortex Agents instead uses the same token-based AI Credit pricing as Agents). One illustrative before/after comparison found a Business Critical customer's monthly Cortex Agents and Cortex Code spend on Claude Sonnet falling from roughly $6,630 to about $2,012 for the same workload — a 70% reduction, driven entirely by the AI Credit decoupling rather than any change in the underlying model's price. Earlier reporting on this repricing described AI Functions as remaining on the older edition-priced currency; Snowflake's current pricing documentation does not support that distinction, and this article treats AI Functions as AI-Credit-priced throughout.

What each vendor bills

Snowflake Cortex (OFFICIAL, docs.snowflake.com/user-guide/snowflake-cortex/pricing and Snowflake's Service Consumption Table, Table 6(a), effective September 30, 2026). AI Functions (AI_COMPLETE, AI_EMBED, AI_CLASSIFY, AI_EXTRACT): confirmed to bill in AI Credits per million tokens, with separate input and output rates per model (both input and output count toward consumption). Current Table 6(a)-consistent dollar rates, at the $2.00 global AI Credit price, include Llama 3.1 8B at $0.24 input / $0.24 output per million tokens, and Claude 4 Opus at $5.00 input / $25.00 output per million — the same dollar figures Claude carries on Anthropic's direct API, Bedrock, and Vertex AI, extending this article's companion managed-LLM-platform finding that Claude's token pricing carries no markup to the Snowflake surface as well. An earlier third-party-reported range of "$0.12 to $5.10 per million tokens" described older or legacy model rates and is superseded by the current Table 6(a) figures used throughout this article. Cortex Agents/Code/Intelligence: billed in the same flat $2.00/$2.20 AI Credit, at the same per-model token rates as AI Functions; one reported example prices Claude Sonnet via Cortex Agents at 1.95 AI Credits per million tokens (blended), or about $3.90 per million at the global rate. Cortex Search: billed per GB of indexed data per month in AI Credits, running continuously regardless of query volume — a real, easy-to-forget always-on cost. Cortex Analyst has two distinct pricing paths depending on how it's invoked: the standalone Cortex Analyst API bills 67 Platform Credits per 1,000 messages — the older, edition-priced currency, not AI Credits, so its dollar cost varies by Snowflake edition — while Cortex Analyst invoked through Cortex Agents or Snowflake CoWork instead uses the same per-token AI Credit pricing as Agents, billing only for tokens actually consumed rather than a flat per-message rate.

Databricks Mosaic AI (OFFICIAL, Databricks Foundation Model Serving pricing pages). Bills in DBUs (Databricks Units, a proprietary compute-normalization metric), reported at roughly $0.07 per DBU for foundation model serving pay-per-token pricing. Per-model DBU-per-million-token rates are published directly by Databricks: Claude Sonnet 4.5: 42.857 DBU/1M input, 214.286 DBU/1M output, which at $0.07/DBU works out to $3.00 input / $15.00 output per million tokens — identical to Anthropic's direct API and Bedrock/Vertex rates, Claude's token pricing in this comparison follows the same no-markup pattern across the modeled surfaces. Llama 3.3 70B: 7.143 DBU/1M input, 21.429 DBU/1M output, working out to $0.50 input / $1.50 output per million — notably, this is cheaper on a blended 5:1 basis (about $667 per billion tokens) than Together AI's direct rate for the same model ($880 per billion), the opposite of the markup pattern found on AWS Bedrock and Google Vertex AI for this same model in this site's companion managed-LLM-platform comparison. Provisioned Throughput (dedicated, reserved capacity) is billed per DBU-hour instead, requiring a minimum concurrency commitment and running 24/7 regardless of actual traffic.

BigQuery AI (OFFICIAL, Google Cloud BigQuery documentation). BigQuery's AI.GENERATE and AI.GENERATE_TEXT functions have no standalone token rate of their own. Google's own documentation states the billing explicitly: BigQuery ML charges apply for the data processed (standard BigQuery bytes-scanned pricing), and a separate Vertex AI charge applies for the actual model call — billed at Vertex's batch API rate for any Gemini 2.0-or-later model, roughly half of Vertex's standard synchronous rate. Partner models (Anthropic Claude, Llama, Mistral) are supported through AI.GENERATE_TEXT specifically and bill at their respective Vertex Model Garden rates, following the same markup pattern documented for open-weight models in this site's companion managed-LLM-platform article.

Cost per 1 billion tokens (5:1 input:output ratio)

Formula: cost per 1B = (5 × input rate + output rate) ÷ 6 × 1,000, or the vendor's own blended per-million rate where that is how it publishes pricing.

The roughly 35x range within Snowflake's own AI Functions rate card (Llama 3.1 8B at $240 versus Claude 4 Opus at $8,333) reflects model choice, not platform markup — the same lesson as this site's companion managed-LLM-platform comparison applies here too: which model you call matters more than which warehouse you call it from. Notably, Snowflake's Claude 4 Opus figure matches the same $5.00/$25.00 per-million dollar rate Claude carries on Anthropic's direct API, Bedrock, and Vertex AI, extending the no-markup pattern for Claude specifically to Snowflake as well.

The always-on cost nobody budgets for: Cortex Search serving

Formula: monthly Cortex Search cost = indexed GB × credits/GB-month × $2.00–$2.20. One reported example priced a 50 GB index at roughly 315 credits a month, or about $630–$693 a month at the new AI Credit rate, billed continuously whether or not anyone queries the index. This is structurally similar to a always-on vector database charge (covered in this site's companion vector-database-cost article) and is easy to omit from a token-based AI budget that focuses only on inference cost.

Sensitivity

1. Which model you call on Snowflake AI Functions. Confirmed from the current Service Consumption Table, this alone spans roughly 35x between Llama 3.1 8B ($240/B) and Claude 4 Opus ($8,333/B).
2. Model choice. The single largest lever on every platform in this article; roughly a 35x spread within Snowflake's own AI Functions rate card alone, from small open-source models to frontier ones.
3. Snowflake edition — no longer a factor for AI Functions. Because AI Functions now bill in the flat, edition-independent AI Credit currency (confirmed by Snowflake's current documentation), a Business Critical or VPS customer pays the same per-token AI Functions rate as a Standard-edition customer; edition only still affects non-AI warehouse compute and storage, billed separately in Platform Credits.
4. Retry behavior on the standalone Cortex Analyst API. Each retried message is a new billable message at 67 Platform Credits per 1,000 — a chatty or error-prone integration multiplies this cost directly, and because this is edition-priced Platform Credits rather than AI Credits, the actual dollar impact also depends on your Snowflake edition.

Budgeting traps

• Budgeting Snowflake Cortex off a pre-April-2026 cost model. The AI Credit decoupling materially changed the math for any customer above Standard edition pricing.
• Assuming Cortex AI Functions still bill in edition-priced Platform Credits. Snowflake's current documentation confirms AI Functions are on the same flat AI Credit currency as Cortex Agents, Code, Intelligence, and Search.
• Forgetting Cortex Search's continuous, indexed-GB-based charge. It runs regardless of query volume, unlike the token-based functions.
• Assuming Databricks marks up every open-weight model the way AWS Bedrock and Google Vertex AI do. At least for Llama 3.3 70B, Databricks' own published DBU rate works out cheaper than a dedicated inference provider's direct price.

What to ask before you buy

Confirm the current per-model rate for your specific model directly against Snowflake's Service Consumption Table (Table 6(a)), since rates are updated over time and this article's examples cover only two representative models. For BigQuery, request a combined estimate covering both the BigQuery compute charge and the separate Vertex AI model-call charge, since neither alone represents the full cost of a text-generation query.