Snowflake Cortex in 2026: what Cortex Analyst and CoWork (formerly Snowflake Intelligence) do and cost, with the arithmetic shown.
Ka Ling Wu
Co-Founder & CEO, Upsolve AI
10 min
Disclosure: This article is published by Upsolve AI. Where our product appears alongside others, the comparison is based on publicly available information as of November 2026.
Snowflake Cortex is Snowflake’s suite of AI features, running inside its platform and billed by consumption. Two 2026 changes make older guides unreliable: Snowflake Intelligence became Snowflake CoWork in June, and in August Snowflake began steering Cortex Analyst users to Cortex Agents, which changes the bill from per message to per token.
The second change matters more to your budget. Under the assumptions below, the same 5,000 questions a month cost $1,005 through the Analyst API and anywhere from $439 to $2,194 through Cortex Agents.
Key Takeaways |
|---|
|
What Is Snowflake Cortex?
Snowflake Cortex, also marketed as Cortex AI, is the umbrella for Snowflake’s LLM-based features. Buyers usually mean one of seven products, which have different users and different meters.
Name today | Earlier name | What it does | Billing, as of November 2026 |
|---|---|---|---|
Cortex AI Functions | Calls models from SQL, such as AI_COMPLETE and AI_CLASSIFY | AI Credits per million tokens | |
Cortex Search | Retrieves from unstructured data | 6.3 AI Credits per GB-month indexed, plus embedding tokens | |
Cortex Analyst | Turns questions into SQL using a semantic view | Per message or per token (see below) | |
Cortex Agents | Plans, calls tools, returns an answer | AI Credits per million tokens | |
Snowflake CoWork | Snowflake Intelligence | Ready-made agent app for business users | AI Credits per million tokens |
Snowflake CoCo | Cortex Code | Coding agent for data teams | AI Credits per million tokens, or a subscription for individual CLI and Desktop users |
Cortex AI Gateway | Governs model and tool access for AI clients | AI Credits per token at each model’s rate |
Rates come from Snowflake’s Service Consumption Table, effective October 29, 2026.
Two of these get mistaken for analytics products. CoCo is for people who write code. Its docs sit under a /cortex-code/ URL, and the pricing page still says Cortex Code. It has its own rate card: Claude Sonnet 4.6 input costs 1.65 AI Credits per million tokens in CoCo and 1.95 in CoWork.
Cortex AI Gateway doesn’t answer business questions. It controls how AI clients reach models and MCP tools, with tracing and spend attribution, and its tool governance is still in preview. It matters if third-party agents are spreading across your company.
What Happened to Snowflake Intelligence?
Snowflake Intelligence is now Snowflake CoWork. Snowflake’s own FAQ says the name changed as the product grew from conversational analytics into a personal work agent, and that existing customers don’t need to do anything.
The product did change. An April 2026 post, still under the old name, had already announced MCP connectors, Skills and Deep Research, so a guide describing a chat box over semantic views is out of date. Under the hood, CoWork still routes each question through the Cortex Agent API, which calls Cortex Analyst for structured data and Cortex Search for documents.
The rename stopped short of the plumbing. Search for the old name in two places:
Setup: the getting-started guide still runs
CREATE SNOWFLAKE INTELLIGENCEand stores agents inSNOWFLAKE_INTELLIGENCE.AGENTS.Finance reporting: the
SNOWFLAKE_COWORK_USAGE_HISTORYview lists its service type asSNOWFLAKE_INTELLIGENCE. A cost query filtering on a CoWork service type would miss it.
Read the preview labels too. Per Snowflake’s November 3 announcement, Snowflake for Slack, the unified mobile app and CoWork Profiles are in public preview, and always-on Bots are “private preview soon”. Cortex Sense, which promises to understand your data without manual semantic-model setup, is also pre-release.
The docs list limits the product page doesn’t. Requests often take longer than a minute, long threads lose earlier context, and Deep Research can take up to 10 minutes. Geographic map charts aren’t supported.
Snowflake Cortex Analyst: Still Works, No Longer the Front Door
Snowflake Cortex Analyst is a REST API that turns a plain-English question into SQL, guided by a semantic view. It has no interface of its own. You build the app.
On August 28, 2026, Snowflake recommended moving to Cortex Agents, saying it supports every Analyst capability with higher answer quality. Existing apps keep working and the REST API remains available, so this isn’t a shutdown notice. Semantic views and verified queries carry over.
The standalone API’s documented limits explain why:
It answers only what SQL can answer. “What trends do you observe?” is out of scope.
It can’t read earlier results. Ask for a product list, then “the revenue of the second product”, and it can’t resolve the reference.
You can’t choose its model. Its preference order starts at Claude Sonnet 4.6 and falls back from there, so results can vary by region and access settings.
Generated queries are capped at 2,048 tokens, and large queries can trigger retries.
Moving to Agents brings a rollout problem that catches teams late. A person querying a semantic view needs SELECT on the view only, but Cortex Agents need SELECT on the view and its underlying tables. If access has lived in service accounts or your BI layer, you now have a grants project, and row access and masking policies on the base tables become your safety net.
Agents also don’t join across separate semantic views, and Snowflake advises staying under about 100,000 tokens per view. Our guide to Snowflake semantic views covers modeling in depth.
The standalone API still makes sense for one case: a narrow text-to-SQL feature where a fixed price per message is worth more than multi-step reasoning.
What Snowflake Cortex Costs: A Worked Example
Snowflake AI features have no per-seat fees. You pay for consumption, and costs add up across every service an agent calls.
Assumptions, which are ours and not measurements: Enterprise Edition on AWS US East, on demand, so a Platform Credit costs $3.00. Global routing, so an AI Credit costs $2.00. 5,000 questions a month. No prompt caching. All tokens priced at Claude Sonnet 4.6 in Table 6(d): 1.95 AI Credits per million input tokens and 9.76 per million output.
Route | Tokens per question | AI Credits per question | Cost per question | At 5,000 a month |
|---|---|---|---|---|
Agents, light | 15,000 in, 1,500 out | 0.0439 | $0.09 | $439 |
Agents, base | 35,000 in, 2,500 out | 0.0927 | $0.19 | $927 |
Agents, heavy (large semantic view, longer thread) | 100,000 in, 2,500 out | 0.2194 | $0.44 | $2,194 |
For the base case: (35,000 × 1.95 + 2,500 × 9.76) ÷ 1,000,000 = 0.09265 AI Credits. Multiply by $2.00 to get $0.1853 a question, then by 5,000 to get $926.50.
The standalone API ignores tokens. At 67 Platform Credits per 1,000 messages, 5,000 messages use 335 credits. That’s $670 on Standard, $1,005 on Enterprise and $1,340 on Business Critical, or $0.134, $0.201 and $0.268 a message at AWS US East on-demand rates.
Warehouse compute sits on top of every row, because the generated SQL runs on your warehouse. If those questions keep a Small warehouse (2 credits an hour) running three hours per business day, that’s 66 hours × 2 × $3.00 = $396. This applies whichever tool writes the SQL.
Don’t trust our token counts. Run 100 real questions, then read TOKENS_GRANULAR in the CORTEX_AGENT_USAGE_HISTORY view, which also shows cache reads. Those cost about a tenth of standard input, so your number could land well below ours.
The per-token meter rewards a small semantic view and short threads. The per-message meter ignores both, which is why the cheaper route depends on your own usage.
The traps that move this number:
Discounts stop at the AI line. The Platform Credit discount doesn’t apply to AI Credits. Contract-size tiers exist, but the AI Credit is still $2.00 below $1.2 million in annual contract value and reaches only $1.88 at $40 million.
Regional routing adds 10%. Global routing costs $2.00 and regional $2.20.
AWS_UScounts as regional, so a data residency setting takes the base case from $927 to $1,019.You choose the orchestration model, not Analyst’s. Snowflake says you can’t directly specify the model for the Analyst tool, only restrict models through access controls, and it advises against that without a compliance need. The pricing page ties the rate to the orchestration model and doesn’t spell out what happens when the two differ, so check your usage data.
Search bills while idle. Cortex Search serving runs continuously while the service is resumed. A 20 GB index is 126 AI Credits, or $252 a month, before anyone asks anything.
Budgets are slow, and standalone Analyst has none. Budget actions can take up to eight hours, or two with the latency-optimized option. Per-user quotas block within minutes but aren’t pooled. Budgets for Cortex Analyst are listed as “not supported nor planned”.
Everyone can spend by default. CORTEX_USER is granted to PUBLIC until you revoke it.
Promotional model prices end. Table 6 flags increases on some models, including 100% on
gemini-4-argonon December 1, 2026. Check before you pin a model.
Where a Governed Data Agent on Snowflake Fits
If all your data is in Snowflake, your users are employees and your team will maintain semantic views, use CoWork. It inherits your row access and masking policies, and there’s no second vendor to review.
Being in Slack or MCP clients is no longer much of a differentiator. CoWork connects to Slack through MCP, and external MCP clients can reach Snowflake agents.
What still separates a dedicated agent, using Upsolve as the example:
Data beyond Snowflake. Generated SQL runs on the Snowflake engine. Snowflake’s new Business Ontology and Horizon Context extend definitions and metadata to other systems, which isn’t the same as answering a question that needs them. Upsolve works across Snowflake, BigQuery, Databricks, Redshift and Postgres at once.
Customer-facing analytics. CoWork is built for your own knowledge workers. To put Cortex in your product, you’d build the interface on the Agents API, where permissions come from the Snowflake context of each request, and map each customer’s users to that context yourself. Upsolve’s embed issues a token per user, and its row rules filter every query by tenant. Unless a rule is marked optional, a missing property makes the query fail instead of returning everything.
Context that isn’t tied to one vendor. Snowflake has verified queries and evaluations, but they live in Snowflake objects. Upsolve’s golden assets double as test cases (tests are in beta), with a trace of each answer, and that context sits outside any one warehouse.
No semantic layer yet. Upsolve doesn’t need one to start. Semantic View Autopilot and Cortex Sense are narrowing this gap, so weigh it less each quarter.
Upsolve’s Team plan, the tier that adds row-level security, RBAC and embedding, costs $2,000 a month with 10,000 credits, or $1,600 a month on annual billing. A standard analytical question uses 5-10 credits.
The case for a dedicated agent is reach: data beyond one warehouse and answers inside your product. It also makes a question’s cost easy to bound: Upsolve caps a question at 20 credits, while on Agents you’d set an orchestration budget to limit a run. Upsolve’s rules live in its own data model, so that’s a second place to maintain access logic. SAML SSO and HIPAA need its Enterprise plan.
For the wider market, see our roundup of AI analytics tools.
Snowflake Cortex vs a Dedicated Data Agent
As of November 2026. Snowflake dollar figures use the assumptions above. Warehouse compute is extra in every column.
Cortex Analyst API | Cortex Agents | Snowflake CoWork | Upsolve Team | |
|---|---|---|---|---|
What it is | Text-to-SQL API | Agent platform you build on | Ready-made agent app | Data agent and context platform |
Where users ask | Your app | Your app, CoWork, CoCo, MCP | Web, iOS, Slack (preview) | Slack, Teams, embedded SDK, MCP |
Pricing | 67 Platform Credits per 1,000 messages | AI Credits per token | AI Credits per token, no seat fee | $2,000/mo, 10,000 credits, $0.05 overage |
5,000 questions a month | $1,005 (Enterprise) | $439-$2,194 | $439-$2,194, same rate table | Plan price, plus overage at 5-10 credits a question |
Semantic layer | Required | Used through Analyst | Used through Analyst | Not required to start |
Data scope | Snowflake | Snowflake, plus tools via MCP | Snowflake, plus tools via MCP | Snowflake, BigQuery, Databricks, Redshift, Postgres |
Access control | Snowflake RBAC | Snowflake RBAC and policies | Snowflake RBAC and policies | Row-level security and RBAC set in Upsolve |
Spend controls | Usage views; no budgets | Budgets, per-user quotas, run budget | Budgets, per-user quotas | Monthly credits, 20-credit cap per question |
Customer-facing embedding | Build it yourself | Build it yourself | Not its purpose | Included, multi-tenant |
Best for | One narrow SQL feature | Custom agents on Snowflake data | Employees, all data in Snowflake | Several warehouses or customer-facing analytics |
Frequently Asked Questions
What is Snowflake Cortex?
Snowflake Cortex is Snowflake’s suite of AI features: AI Functions, Cortex Search, Cortex Analyst, Cortex Agents, Snowflake CoWork and more. They run inside Snowflake and bill by consumption.
Is Snowflake Intelligence the same as Snowflake CoWork?
Yes. Snowflake announced the new name at Summit on June 2, 2026, and says existing customers need to take no action. Some SQL commands and billing labels still use the old name.
How much does Snowflake Cortex cost?
There’s no per-seat fee, so it depends on the feature and tokens used. AI Credits cost $2.00 or $2.20, and token rates vary by model. Under our labelled assumptions, 5,000 agent questions cost $439-$2,194 a month plus warehouse compute.
Is Snowflake Cortex Analyst being deprecated?
No retirement date appears in Snowflake’s August 2026 note. Snowflake recommends moving to Cortex Agents, and the REST API remains available. Analyst remains the tool Agents use for structured data.
Can Snowflake Cortex query data outside Snowflake?
Analyst’s SQL runs on Snowflake. Agents and CoWork can call outside systems through MCP connectors, which is different from running one governed query across warehouses. If your questions routinely span platforms, test that case first.
If your data lives in more than one warehouse, or the agent has to serve your customers, see Upsolve’s pricing or book a demo.
Sources
Snowflake - Press release, Snowflake CoWork at Summit 26, June 2, 2026. https://www.snowflake.com/en/news/press-releases/snowflake-cowork-powers-the-agentic-enterprise-as-the-personal-agent-for-knowledge-workers-to-work-smarter/
Snowflake Docs - Snowflake AI pricing. https://docs.snowflake.com/en/user-guide/snowflake-cortex/pricing
Snowflake - Snowflake Cortex AI product page. https://www.snowflake.com/en/product/features/cortex/
Snowflake - Service Consumption Table, effective October 29, 2026. https://www.snowflake.com/legal-files/CreditConsumptionTable.pdf
Snowflake Docs - Overview of Snowflake CoCo. https://docs.snowflake.com/en/user-guide/cortex-code/cortex-code
Snowflake Docs - Cortex AI Gateway. https://docs.snowflake.com/en/user-guide/snowflake-cortex/cortex-ai-gateway
Snowflake - Snowflake CoWork product page and FAQ. https://www.snowflake.com/en/product/snowflake-cowork/
Snowflake - From Answers to Action with Your Personal Work Agent, April 21, 2026. https://www.snowflake.com/en/blog/snowflake-intelligence-work-agent/
Snowflake Docs - Overview of Snowflake CoWork. https://docs.snowflake.com/en/user-guide/snowflake-cortex/snowflake-cowork
Snowflake Docs - Getting started with Snowflake CoWork. https://docs.snowflake.com/en/user-guide/snowflake-cortex/snowflake-cowork/getting-started
Snowflake Docs - AI cost management and governance. https://docs.snowflake.com/en/user-guide/snowflake-cortex/governance-and-availability/ai-cost-management-and-governance
Snowflake - Press release, CoCo and CoWork at Expedition 2026, November 3, 2026. https://www.snowflake.com/en/news/press-releases/expedition-2026-coco-cowork/
Snowflake Docs - Snowflake CoWork troubleshooting. https://docs.snowflake.com/en/user-guide/snowflake-cortex/snowflake-cowork/troubleshooting
Snowflake Docs - Cortex Analyst. https://docs.snowflake.com/en/user-guide/snowflake-cortex/cortex-analyst
Snowflake Docs - Release note, transitioning from Cortex Analyst to Cortex Agents, August 28, 2026. https://docs.snowflake.com/en/release-notes/2026/other/2026-08-28-cortex-analyst-transition-cortex-agents
Snowflake Docs - Best practices for developing and deploying semantic views. https://docs.snowflake.com/en/user-guide/views-semantic/best-practices-dev
Snowflake Docs - Best practices for modeling semantic views. https://docs.snowflake.com/en/user-guide/views-semantic/best-practices-modeling
Snowflake Docs - Integrate tools and data with Snowflake CoWork. https://docs.snowflake.com/en/user-guide/snowflake-cortex/snowflake-cowork/integrate-tools
Snowflake - Press release, Business Ontology and Cortex AI Gateway, November 3, 2026. https://www.snowflake.com/en/news/press-releases/expedition-2026-business-ontology-ai-gateway/
Snowflake Docs - Cortex Agents. https://docs.snowflake.com/en/user-guide/snowflake-cortex/cortex-agents
Upsolve AI - Company and product summary. https://upsolve.ai/
Upsolve AI - Pricing. https://upsolve.ai/pricing
Upsolve Docs - Embed Upsolve in your product. https://docs.upsolve.ai/embedding
Upsolve Docs - Data Governance. https://docs.upsolve.ai/ai-agent-builder/data-governance
Upsolve Docs - Golden Assets. https://docs.upsolve.ai/ai-agent-builder/golden-assets
Upsolve Docs - Agent Observability. https://docs.upsolve.ai/ai-agent-builder/agent-observability
Upsolve AI - Snowflake Semantic Views: The 2026 Guide. https://upsolve.ai/blog/snowflake-semantic-layer
Try Upsolve for Embedded Dashboards & AI Insights
Embed dashboards and AI insights directly into your product, with no heavy engineering required.
Fast setup
Built for SaaS products
30‑day free trial
