What is CRE pro forma building? CRE pro forma building is the workflow of constructing a multi-year operating projection for a commercial property starting from broker materials (offering memorandum, T12 operating statement, rent roll), defining tab structure, populating revenue and expense growth assumptions, layering loan amortization, and producing IRR and equity multiple outputs for the investment committee. This article is a workflow comparison rather than a benchmark, focused on how to actually build the pro forma using Claude Opus 4.7 versus Gemini 3.1 Pro, including the prompts that work and the validation steps each model requires. For the benchmark-style head-to-head on financial modeling, see our Claude Opus 4.7 vs Gemini 3 financial modeling guide.
Key Takeaways
- Claude Opus 4.7 produces more reliable IC ready pro formas because its xhigh effort level surfaces logical inconsistencies (e.g., expense growth that exceeds revenue growth) before the model is finalized.
- Gemini 3.1 Pro builds faster and at roughly 60% lower cost ($2 input and $12 output per million tokens versus Claude's $5 and $25), making it the better choice for high-volume preliminary screening.
- The pro forma BUILDING workflow benefits from a tab-by-tab prompting structure rather than a single shot prompt. Both models perform meaningfully better with this approach.
- Claude's task budget feature is the killer feature for long pro forma builds because it prevents the model from running out of tokens halfway through a complex five tab build.
- Neither model should set the underwriting assumptions (rent growth, expense growth, exit cap) without sponsor review against actual market data. Both will produce defensible-looking but unsupported numbers if not anchored.
The Pro Forma Building Workflow
A production-grade CRE pro forma is built in five tabs: Assumptions, Revenue, Operating Expenses, Cash Flow Summary, and Returns. Each tab depends on the prior tab. Building this from scratch using AI requires a tab-by-tab prompt sequence rather than asking for the whole model at once. The reason is that long-output prompts hit attention budget limits and the model starts copying boilerplate from earlier examples in its training data instead of computing from the inputs.
For both Claude Opus 4.7 and Gemini 3.1 Pro, the workflow we use with clients is: (1) extract the inputs from the broker package; (2) build the Assumptions tab with sponsor review; (3) build the Revenue tab; (4) build the Operating Expenses tab; (5) build Cash Flow Summary with loan amortization; (6) build Returns with IRR sensitivity. Each step is a separate prompt that references the prior step's output.
Step 1: Extracting Inputs From the Broker Package
The first prompt extracts unit count, rent roll summary, T12 NOI, asking price, and any sponsor-stated assumptions from the OM. We tested both models on a 184 unit garden multifamily OM with a T12 statement and rent roll.
Claude Opus 4.7: Extracted 100% of the structured inputs accurately. Flagged that the broker's stated cap rate (5.85%) used a forward NOI that assumed full lease-up, while the in-place cap rate (using T12 NOI of $1.62 million on the $32.5 million ask) was 4.99%. This discrepancy is the kind of thing Claude consistently catches that other models do not.
Gemini 3.1 Pro: Extracted 100% of the structured inputs and produced the table cleanly. Did not flag the in-place versus forward cap rate distinction unprompted, but caught it on a follow-up question. The 1 million plus token context window let Gemini ingest the full OM, T12, and rent roll in a single prompt with room for instructions.
Step 2: Building the Assumptions Tab
The Assumptions tab is the foundation. It defines hold period, year one revenue growth, ongoing revenue growth, expense growth, exit cap rate, and loan structure. Sponsors who skip this step or allow the AI to set defaults end up with garbage models.
Recommended prompt structure: "Using the inputs above, build an Assumptions tab for a five year hold. I want to specify each assumption myself. Show me the standard ranges for a stabilized class B multifamily in [submarket], cite where each range comes from, then leave blanks for me to fill in."
Claude Opus 4.7: Produced a clean Assumptions tab with cited ranges. Year one rent growth range cited as 2.5 to 4.5% based on submarket trailing data. Expense growth range cited as 3.5 to 4.5% based on inflation plus property tax reassessment risk. Exit cap rate cited as 25 to 75 basis points above going-in cap rate. Used the xhigh effort level which surfaces these reasoning steps explicitly.
Gemini 3.1 Pro: Produced a similar Assumptions tab with comparable ranges. Cited the ranges with slightly less specificity but accurately. Faster output (38 seconds versus Claude's 1 minute 24 seconds at xhigh effort).
Step 3: Building the Revenue Tab
The Revenue tab projects unit-level rent growth, occupancy, other income (parking, storage, laundry), and any planned rent bumps post-renovation. This is where most pro forma models go wrong because the AI defaults to blanket rent growth assumptions instead of unit-by-unit modeling.
Claude Opus 4.7: Built a unit-by-unit revenue tab with explicit handling of below-market units (rent bumps to market on lease rollover), classic-style apartments versus renovated apartments (different rent growth trajectories), and seasonal occupancy variation. The output included a year-by-year rent roll waterfall showing how unit-level growth aggregates to project NOI.
Gemini 3.1 Pro: Built a similar revenue tab with cleaner table formatting (Gemini's multimodal architecture handles structured data well) but defaulted to a blanket 3.0% rent growth across all units unless specifically prompted to differentiate. Required a follow-up prompt to model the below-market lease rollover dynamics.
Step 4: Building Operating Expenses
OpEx modeling is where AI tools tend to be the weakest because expense growth is non-uniform: property taxes step up at reassessment, insurance has been growing 12 to 18% annually in select markets per NMHC research, payroll grows with wage inflation, and utilities depend on weather and tenant submetering.
Claude Opus 4.7: Modeled each expense line separately with its own growth rate. Property tax line included a step-up in year two based on the assumed reassessment timing. Insurance line carried a 14% year one growth assumption with a tapering curve. The xhigh effort level made the differential growth explicit rather than buried.
Gemini 3.1 Pro: Built a comparable OpEx tab. Defaulted to a single 3.5% blanket expense growth unless prompted to differentiate. After the prompt to differentiate, output was comparable to Claude's. For more on the multifamily-specific underwriting context, see our AI multifamily value-add underwriting guide.
Step 5: Cash Flow Summary With Loan Amortization
The Cash Flow Summary tab combines NOI from prior tabs, layers debt service, calculates DSCR by year, and produces cash flow before tax. This is where Claude's task budget feature matters: by year five of a complex model, output token consumption is high, and a model that hits its attention limit produces incomplete or hallucinated outputs.
Claude Opus 4.7: Used the task budget set to 25,000 tokens for the build. Completed the full five year cash flow summary with loan amortization (interest only year one, principal and interest years two through five) cleanly. DSCR by year calculated correctly.
Gemini 3.1 Pro: Completed the same build faster (1 minute 47 seconds versus Claude's 3 minutes 12 seconds) but on year five the model truncated the amortization schedule by three months, producing a marginally incorrect ending principal balance. This required a follow-up prompt to extend.
Step 6: Returns and IRR Sensitivity
The Returns tab calculates project-level IRR, equity multiple, average cash on cash, and an exit cap sensitivity grid. Both models handle this competently when given clean inputs from prior tabs.
Claude Opus 4.7: Built a 5x5 IRR sensitivity grid (exit cap rate from 5.50% to 6.50% in 25bp steps, year one rent growth from 1.5% to 4.5% in 75bp steps). All cells calculated correctly. Equity multiple of 1.78x at base case with project IRR of 14.2%.
Gemini 3.1 Pro: Built the same 5x5 grid faster, all cells correct. Project IRR matched at 14.2% and equity multiple at 1.78x. The math is identical when inputs are identical.
Pricing Comparison for Pro Forma Volume
For a sponsor building 20 pro formas per month, average input around 60,000 tokens per build (broker materials plus prior tab outputs), average output around 12,000 tokens (full five tab model with explanations):
- Claude Opus 4.7: 1.2M input at $5/M = $6.00; 240K output at $25/M = $6.00. Total: $12.00 per month, $0.60 per pro forma.
- Gemini 3.1 Pro: 1.2M at $2/M = $2.40; 240K at $12/M = $2.88. Total: $5.28 per month, $0.26 per pro forma.
For high volume preliminary screening (50 plus pro formas per month for an institutional acquirer), the cost gap matters. For ten or fewer per month at IC quality, the partner-hour savings dominate the API cost difference, and Claude's higher accuracy on differentiated growth assumptions is worth the premium. CRE professionals looking to operationalize this kind of pro forma workflow can connect with The AI Consulting Network for hands-on implementation support.
Which Model Should Sponsors Choose?
For institutional sponsors building IC ready pro formas where every assumption needs to be defended in front of an LP, Claude Opus 4.7 is the safer choice. The xhigh effort level surfaces reasoning that LPs will ask about. For high volume preliminary screening on twenty plus deals per month before LOI, Gemini 3.1 Pro is the cost leader and produces functionally equivalent output when prompted carefully. CRE investors looking to build this kind of workflow into their underwriting process can reach out to Avi Hacker, J.D. at The AI Consulting Network.
Frequently Asked Questions
Q: Should I build the pro forma in Excel or in chat output?
A: For institutional capital, build in Excel. The AI tool generates the structure and the formulas, but the final model needs to be a working spreadsheet that LPs can audit. Both Claude and Gemini can output Excel-compatible formula structures; the chat output is for review, not for the LP package.
Q: Why does the tab-by-tab prompt approach work better than a single shot?
A: Long-output prompts cause both models to drift toward generic templates rather than computing from your specific inputs. By prompting tab-by-tab and feeding the prior tab's output as context, you keep each prompt within the attention window where the model is sharpest.
Q: Can either model handle a value-add pro forma with renovation capex?
A: Yes. Both Claude and Gemini handle renovation capex, post-renovation rent bumps, and the lease rollover schedule that drives the value-add thesis. Claude's xhigh effort level produces tighter logic on the renovation timing-to-rent-bump connection.
Q: How do I validate the pro forma without rebuilding it manually?
A: Run a sanity check: project NOI year three should be roughly equal to T12 NOI grown by your assumed rent growth less your assumed expense growth, weighted by the revenue versus expense base. If the model output is more than 5% off this back-of-envelope math, something is wrong in the build.
Q: Does Gemini's 2 million token context window matter for pro forma building?
A: Only for institutional acquirers with very large data rooms. For a single property pro forma, even a complex one, you will not exceed Claude's 1 million token window. The context window difference matters more for portfolio underwriting than for single-asset pro forma builds.