When To Use Which Mode In Snowflake CoWork - Chat vs Deep Research vs Extended Thinking

Snowflake renamed Snowflake Intelligence to Snowflake CoWork and added Deep Research and Extended Thinking modes, each serving distinct use cases. Standard Chat balances speed and quality for quick queries, Extended Thinking boosts reasoning on a single answer for complex questions, and Deep Research decomposes questions into parallel sub-investigations across structured and unstructured data to produce a synthesized report. Choosing the wrong mode wastes time or credits, according to the Snowflake documentation.

In this blog we are going to discuss the different interaction modes available inside Snowflake CoWork and their specific use case. I will try to make the use case as clear as possible and hope after reading this blog you can easily decide which mode to reach for based on your requirement. I am also going to include a brief of each one, so you are aware of what we are talking about. If you have been on Summit calls or reading the release notes lately, you already know Snowflake renamed Snowflake Intelligence to Snowflake CoWork and packed in a bunch of new capabilities around it. Two of those capabilities, Deep Research and Extended Thinking, get confused with each other a lot because both sound like “make the AI think more.” They are not the same thing and picking the wrong one either wastes your time or wastes your credits. Let’s go through each. Standard Chat the default mode — This is what you get out of the box the moment you open Snowflake CoWork and ask a question. Snowflake CoWork uses AI powered data agents that understand your question, perform analysis, generate insights, and can take action, all through the Cortex Agent API sitting behind the scenes. By default, the agent balances speed and quality automatically, so for most day to day questions you don’t need to touch any toggle at all. It works over both structured and unstructured enterprise data, meaning your tables and your documents, at the same time. Results come back as natural language, tables, or auto generated charts depending on what makes sense for the question. Trends render as charts, lookups render as tables. Every answer respects your existing row access policies and column level security, since CoWork inherits Snowflake’s governance model automatically. It doesn’t bypass anything you already have locked down. You can save any generated chart or table as an Artifact, which is basically a persistent object you can revisit or share without regenerating it every time. Use-Case: Quick metric lookups, for example “how are Q4 sales trending.” Interactive back and forth where you are drilling into details one question at a time. Day to day business questions where you don’t need a full investigation, just an answer. Extended Thinking — Extended Thinking is an internal reasoning boost for a single answer. The agent thinks harder before responding, validating the SQL logic, the joins, and the access controls before it actually executes anything against your data. It is toggled on right inside the chat window, and once you turn it on, it stays selected for the session. It does not go explore multiple data domains or run parallel investigations, it stays focused on getting one answer right by double checking its own reasoning before committing to a query. Because it is doing more validation work behind the scenes, it takes more time and consumes more tokens compared to a standard chat response. This is the same idea used on the model side with Claude on Cortex, where models like Claude Sonnet can run in extended thinking mode and self reflect before answering, giving Snowflake CoWork this reasoning boost to draw on. Use-Case: Complex questions where a wrong join or a misread semantic model would give you a confidently wrong answer. Whenever you feel the agent is exploring too many options or you specifically want it to be more thorough on one query. Situations where correctness on a single query matters more than speed, but you don’t need to go pull in unstructured data or external context. Deep Research — Deep Research is a completely different mode. Instead of answering one question directly, it decomposes your question into multiple sub investigations and runs them in parallel across both your structured and unstructured data. Once all the sub investigations are done, it synthesizes everything into one structured report, and every single claim in that report traces back to the source data and the query that produced it. You start it by hitting the + button in the message bar and selecting Deep Research instead of typing directly into standard chat. Investigations can take up to 10 minutes to complete, so this is not something you fire off while waiting for an answer in the next 10 seconds. Once the report is generated, it stays in context for the rest of that thread. Follow up questions like “break that down by region” run as normal chat turns after that, they don’t restart the whole investigation. Deep Research and Extended Thinking are not mutually exclusive, you can actually enable Extended Thinking during a Deep Research investigation if you want deeper reasoning combined with wider exploration. Use-Case: Questions that need a “why,” for example “why has forecast accuracy been declining, and what’s driving the variance.” Anything where the answer likely spans multiple tables or business domains and a single query won’t cut it. When you need a citable report you can actually hand to stakeholders, not just a chat answer that disappears once the thread scrolls away. Identifying patterns, for example “which customer segments show early churn signals, and what do they share.” Not the right choice for quick lookups or single metric questions, standard chat is faster and better suited there. Pros & Cons at a glance — Standard Chat Pros: fastest response, works for the vast majority of daily questions, zero setup or extra clicks needed. Cons: not built for deep multi domain investigations, can occasionally take a shortcut on a complex query that needed more validation. Extended Thinking Pros: catches logic errors before execution, more thorough validation on SQL, joins, and access checks, easy toggle right in the chat window. Cons: slower than standard chat, uses more tokens, still limited to answering a single question rather than exploring broadly. Deep Research Pros: parallel sub investigations across structured and unstructured data, fully cited and traceable report, great for stakeholder ready deliverables, can be combined with Extended Thinking for even deeper analysis. Cons: can take up to 10 minutes per investigation, heaviest on token and credit consumption of the three, overkill for anything that is really just a quick lookup. Cost & things to be aware of — Snowflake has moved to a token based AI Credit model that governs Cortex AI Functions, the REST API, Snowflake CoWork, and Cortex Agents all together, separate from your regular Platform Credits used for warehouse compute. AI Credits are priced the same regardless of your Snowflake edition, currently around $2.00 for Global and a bit more for Regional, unlike Platform Credits which vary by edition. Snowflake CoWork itself is billed in AI Credits based on token consumption since it uses agents to orchestrate multi step workflows behind every question you ask. Deep Research will naturally cost more per investigation than a standard chat answer, since it is running multiple sub investigations in parallel instead of a single query, and each of those sub investigations consumes its own share of tokens. Extended Thinking adds cost on top of a standard answer too, but nowhere near what a full Deep Research investigation runs, since it is still just one question being reasoned through more carefully. Model choice matters a lot here, frontier models cost meaningfully more per token than smaller or open source options, and the gap compounds fast on heavier workloads like Deep Research where many tool calls stack up. Keep an eye on your CORTEX FUNCTIONS QUERY USAGE HISTORY and related usage views if you want to track exactly where your AI Credits are going, especially once teams start leaning on Deep Research regularly for reporting. Some of these features, including Deep Research and parts of Extended Thinking’s underlying capabilities, are still rolling out through public preview, so double check availability on your account before you build a workflow that depends on them. Hope this blog helps you get some clarity on when standard chat is enough, when Extended Thinking is worth the extra tokens, and when the question really needs a full Deep Research investigation instead. Feel free to drop a question in the comment section if you have any doubts regarding this. Give a clap if you like the blog. Stay connected to see many more such cool stuff. Thanks for your support. Disclaimer: Please note opinions expressed in this article are solely my own and do not represent the views or opinions of my employer.