GPT for Sheets: What It Does (and Where It Breaks) (2026)

What GPT for Sheets Functions Actually Do

The core mechanic is a custom function that sends cell content to an LLM and returns the result as a cell value. The Talarian add-on (as of Q2 2026) ships these primary functions:

• =GPT(prompt, [value], [temperature], [max_tokens]) - free-form prompt with optional cell input
• =GPT_LIST(prompt, [value]) - returns a bulleted list
• =GPT_TABLE(prompt, [headers], [value]) - returns structured tabular output
• =GPT_EXTRACT(prompt, value) - pulls specific data from unstructured text
• =GPT_CLASSIFY(value, categories) - categorizes input against a fixed list
• =GPT_TRANSLATE(value, language) - translates content
• =GPT_FORMAT(value, format) - reformats text output

Under the hood, each call goes to OpenAI's API on your own key. You choose the model: gpt-4o runs at $2.50 per million input tokens and $10.00 per million output tokens (OpenAI pricing, May 2026); gpt-4o-mini drops that to $0.15 input / $0.60 output. For most formula-level tasks, mini is plenty.

The add-on also includes a sidebar where you can run one-off prompts against sheet data without writing formulas at all.

Where Cell-Level AI Breaks for Financial Models

The 80% accuracy rate you'll see cited for GPT-based extraction tasks sounds acceptable until you remember that a DCF with one wrong input compounds through 8 linked tabs. Cell-level AI has 3 structural problems that matter in finance.

No model awareness. =GPT_EXTRACT(...) in cell D14 on your Revenue tab doesn't know that F14 on your Returns Analysis tab is dividing by that number. The function returns what looks right in isolation. Whether it ties out is your problem.

Recalculation is destructive. Every time the sheet recalculates - or you open it on a different machine - every GPT cell fires again. Your WACC on one open might be 9.2%; the next might be 9.4%. The Talarian documentation (Q2 2026) states this explicitly: "GPT functions are non-deterministic and will return different results each time they are calculated." You can copy-paste values over formulas to freeze them, but now you've broken your audit trail.

Token costs at scale are real. A 2,000-SKU contribution margin analysis with =GPT_CLASSIFY(A2, categories) dragged down column A sends 2,000 API calls. At gpt-4o pricing, a detailed classification prompt might run 500 input tokens per call - that's 1 million tokens, $2.50, per classification pass. Run it 5 times while testing prompts and you've spent $12.50 on one column. Active monthly use typically runs $15-30/month for moderate workloads.

Where GPT for Sheets Tools Are Actually Useful

The cell-level model works well for exactly one class of problem: transforming unstructured text at the row level, where each row is independent and approximate accuracy is acceptable.

Real use cases that hold up:

CIM parsing. You've extracted text paragraphs from a 40-page CIM into column A and want a first-pass classification by topic. =GPT_CLASSIFY(A2, "revenue drivers, management, risk factors, capex, other") dragged down 200 rows gives you a sortable column in under 2 minutes. You're going to read every paragraph anyway - this just sorts them first.

Company description normalization. Building a comps table with 30 companies from different sources and business descriptions are inconsistent. =GPT("Summarize in one sentence under 20 words: "&A2) standardizes them for the board deck.

Deal note tagging. 500 rows of CRM notes from a BD process. =GPT_CLASSIFY(B2, "pricing discussion, term sheet, due diligence, pass, active") gives your team a pipeline view without manual triage.

None of these are in your live financial model. That's the pattern. GPT functions belong in staging tabs - data prep, not model inputs.

One Non-Obvious GPT for Sheets Problem Worth Knowing

The most common silent failure is numeric extraction from financial statements.

Say you're pulling EBITDA figures from earnings text into a DCF. =GPT_EXTRACT("Extract EBITDA in millions", A2) returns "142.3" - looks clean. But 30% of the time it's grabbing the wrong figure (adjusted vs. reported, trailing vs. forward), and you won't catch it unless you spot-check every row. On a 50-company screening model, that's up to 15 wrong inputs. At a 14.2x EV/EBITDA multiple, a $1M EBITDA error translates to $14.2M in enterprise value. The $420K variance flagged in a Q3 board pack can trace back to exactly this.

The fix is simple but tedious: never route GPT extraction directly into a model input. Always add a validation column that cross-checks the extracted number against a formula-based figure, and flag anything that deviates more than 5%.

GPT for Sheets vs. Native AI in Sheets vs. ModelMonkey

ModelMonkey sits in a different category because it doesn't operate at the cell function level. It reads your full workbook context and can work across tabs. When you ask it to "reconcile the FCFF bridge on the Returns tab against the Cash Flow statement," it knows what's on both tabs. A GPT formula function never sees past the cell it lives in.

Try ModelMonkey free for 14 days - it works in both Google Sheets and Excel.