ChatGPT vs C
ChatGPT vs Claude for Data Analysis: Excel and SQL Handling Compared
An afternoon test plan compares ChatGPT and Claude on Excel and SQL tasks using vendor-documented limits and your own correctness log, without claiming a quality winner.
Sources checked 2 Oct 2026
As read on October 1, 2026, OpenAI’s File Uploads FAQ says ChatGPT accepts spreadsheet uploads on Free and paid plans, subject to plan limits and settings; Anthropic’s upload-files page lists XLSX support and requires code execution and file creation. These pages settle setup, not which assistant writes the more correct Excel formula or SQL query. Chat Picker has not tested either for data analysis, so this is a test plan, not a ranking.
Know these limits before you start
For ChatGPT, OpenAI’s File Uploads FAQ, read on October 1, 2026, caps CSV files and spreadsheets at approximately 50MB, depending on row size. Free users have 3 uploads per day; limits may fall during peak hours. Projects hold 5 files on Free and 25 on Go or Plus. Use a workbook that fits both assistants’ project systems, and log any rejection.
For Claude, Anthropic’s upload-files page, read on October 1, 2026, requires code execution and file creation for XLSX. It sets a 500MB limit per chat file and 30MB per project file; extracted content may add token limits. For non-PDF documents, it extracts text only and does not interpret embedded images.
Data handling is only partly settled. OpenAI’s File Uploads FAQ says chats remain until deletion, files follow chat retention, and files uploaded to Advanced Data Analysis are deleted after a plan-dependent duration. Anthropic’s upload-files page states no chat-file retention period. Check current account controls before using work data.
OpenAI’s accuracy note, read on October 1, 2026, says ChatGPT may be wrong while sounding confident; verify data, quotations, technical information, and references. Claude’s incorrect-response note, read on the same date, warns that Claude can produce convincing but ungrounded quotations and should not be your only source of truth.
For sensitive workflows, OpenAI’s Usage Policy prohibits unauthorized aggregation or distribution of private or sensitive information and high-stakes automated decisions without human review. Anthropic’s Usage Policy calls for relevant human expertise in elevated-risk uses and qualified review of covered advice, recommendations, and subjective decisions. Both were read on October 1, 2026; treat human review as a gate.
The test plan
Run both assistants in the same browser and network where possible. Record each account’s plan and visible settings; run SQL against the same local database copy and indexes.
-
Excel formulas. Input: an XLSX with
order_id,customer_name,region,order_date,units,unit_price,discount, andreturn_flag. Add a blank price, text date, duplicate spelling, and returned order. Prompt: “Create an Excel formula for net revenue by region and month. Use structured references, make it safe to copy down, explain blanks, text dates, duplicate labels, and returns, and do not overwrite the source.” Inspect ranges, separators, copied behavior, and whether the explanation matches. Log the response, edits, errors, and an independently checked result. -
SQL joins and subqueries. Input:
customerswithcustomer_idandregion;orderswithorder_id,customer_id,status, andorder_date; andorder_itemswithorder_id,line_item_id,quantity, andunit_price. Include a null customer key, canceled orders, repeated line-item IDs, and tied timestamps. Prompt: “Write PostgreSQL queries that return each customer’s completed-order revenue and latest order date for the most recent calendar month, with region included. Provide a CTE version and a correlated-subquery version. State row grain, each join’s purpose, and treatment of null, canceled, repeated, and tied rows.” Run both on the same data and inspect execution plans and elapsed times. Log syntax fixes, duplicated or missing rows, assumptions, and plan changes. Timing describes these queries and this dataset, not assistant speed. -
Cleaning and transformation. Input: a CSV with leading and trailing spaces, mixed-case IDs, currency symbols, mixed date formats, blank required fields, and possible duplicate records. Prompt: “Create a reproducible cleaning workflow for this CSV. Work on a copy, show each transformation in order, preserve an audit field for every changed or rejected row, and list all assumptions. Do not silently discard or merge ambiguous records.” Inspect parsing order, destructive steps, auditability, and ambiguous-value handling. Log changed and rejected rows, code edits, and anything you cannot classify safely.
-
Aggregation and pivot summaries. Input: a transaction table with
region,product,channel,order_date,quantity,unit_price,return_flag, andtransaction_id. Include a blank category, repeated transaction ID, and returns. Prompt: “Create a pivot-style summary by region and product and a second summary by month and channel. State filters, aggregation grain, blank handling, and treatment of duplicate transaction IDs and returns. Provide Excel formulas and PostgreSQL SQL, plus a reconciliation check against source totals.” Inspect whether definitions remain consistent and whether the checks expose discrepancies. Log manual adjustments and any mismatch among Excel, SQL, and source totals. -
Errors and edge cases. Input: broken SQL and workbook cases involving a reserved-word alias, missing parenthesis, circular reference, text in a numeric column, division by zero, and a null date. Prompt: “Identify the likely failure in each SQL or Excel input. Return corrected SQL or formulas, explain each failure, and state one assumption that could make a correction wrong. Do not claim to have executed anything.” Inspect which problems are identified, missed, or misdiagnosed. Log every correction and the evidence used to verify it.
How to read your results
Run each task in a fresh chat, then repeat any result you rely on with the same file and wording. Save the first response before editing it.
Use hard gates for non-executable syntax, wrong row grain, duplicated or missing values, silent data loss, or explanations that conflict with the formula or query. Then compare agreement with an independently prepared answer, manual corrections, clarity of assumptions, reproducibility, and edge-case handling. Keep each response beside its log entry.
If both assistants pass your must-have checks, decide from your workflow and documented plan limits. If neither passes, fix the inputs or narrow the task before comparing again. One afternoon’s result supports that test, not a general winner.
Where the plans differ
Use Chat Picker’s Claude vs ChatGPT matrix and check the rows for Free plan, Low-cost tier, Main paid plan, Heavy-use plans, Context window in the app, Usage limits, File/project limits, Training on your chats, Ads, and Team plan. Start with file and project limits, account settings, usage limits, and context window; then check the commercial rows before subscribing.
Each matrix figure links to its vendor source and read date, while unconfirmed cells are marked “Not verified” and left blank. Confirm the vendor page before purchase because plans and regional prices change.