What is a real estate investor underwriting memo? A real estate investor underwriting memo is the one to three page document a sponsor or analyst writes to argue for or against a deal in front of an investment committee. The right AI for this workflow in May 2026 depends on whether you optimize for accuracy (Claude Opus 4.7), spreadsheet integration and speed (ChatGPT GPT-5.5), or cost and real-time market context (Grok 4.3). This is the first three-way head-to-head we have run that adds Grok to the usual ChatGPT vs Claude conversation, and the difference matters for CRE shops that produce more than ten memos per month. For comprehensive coverage of the full underwriting stack, see our AI tools for real estate investors guide.
Key Takeaways
- Claude Opus 4.7 wins on memo writing quality, broker fluff detection, and clean three-page IC formatting that survives senior partner review without rewrites.
- ChatGPT GPT-5.5 wins on workflow integration, producing the memo and the supporting Excel pro forma in a single pass via the ChatGPT for Excel add-in.
- Grok 4.3 wins on cost ($1.25 input and $2.50 output per million tokens) and on real-time X data for sentiment around a market or sponsor.
- For ten plus memos per month, the lowest total cost path is Grok 4.3 for first-draft generation, then Claude Opus 4.7 for IC ready polish.
- None of the three should write the financial assumptions without sponsor review. All three still produce confident-sounding errors on rent growth, cap rate selection, and exit assumptions.
The Three Models in May 2026
Each model has shifted enough in the last sixty days that prior comparisons are out of date. Here is what is current as of May 2026.
Claude Opus 4.7 released April 16, 2026. It supports a 1 million token context window, 128,000 max output tokens, adaptive thinking, and a new xhigh effort level for the hardest tasks. Pricing held at $5 per million input tokens and $25 per million output tokens, though the new tokenizer can produce up to 35% more tokens for the same input. The model added high-resolution image support up to 2,576 pixels and a task budget feature that gives Claude a running token countdown during long agentic tasks.
ChatGPT GPT-5.5 Instant replaced GPT-5.3 Instant as the default ChatGPT model on May 5, 2026, two days before this article. It carries forward GPT-5.4's investment banking benchmark gains (87.3% on a three-statement model build versus 43.7% for the prior generation), the 33% reduction in factual errors versus GPT-5.2, and the ChatGPT for Excel add-in. ChatGPT for Excel and Google Sheets is now generally available across all paid plans.
Grok 4.3 launched in beta on April 17, 2026. xAI priced it at $1.25 per million input tokens and $2.50 per million output tokens, roughly 5x cheaper than Claude Opus 4.7 on input and 10x cheaper on output. Grok 4.3 supports a 2 million token context window, real-time access to X (formerly Twitter) data, and SOC 2 Type II auditing for enterprise compliance. It scored 91.7% on AIME 2025 mathematics and reduced hallucination rates roughly 3x versus Grok 4.
Test 1: First-Draft IC Memo From a Broker Package
We fed each model the same 64 page multifamily offering memorandum (a stabilized 230 unit garden-style asset in Phoenix), a T12 operating statement, a current rent roll, and the sponsor's instructions: write a three page IC memo with a recommendation, key risks, and an underwriting summary table.
Claude Opus 4.7 produced a three page memo in 4 minutes 12 seconds with the cleanest formatting. It correctly flagged that the broker's projected rent growth assumed 4.8% year one when comparable submarket data supports 2.9%, and it included a reconciliation paragraph showing the impact on year three NOI. The memo opened with a one paragraph thesis, used numbered risk callouts, and closed with a recommendation. No formatting cleanup needed.
ChatGPT GPT-5.5 produced a three page memo in 3 minutes 38 seconds. The prose was tighter than Claude's in places, but it accepted the broker's rent growth assumption without challenging it. It also produced a parallel Excel pro forma showing the underwriting math (the ChatGPT for Excel integration is the key differentiator here). For shops that want the memo and the model in a single pass, this is a real workflow advantage.
Grok 4.3 produced a three page memo in 2 minutes 51 seconds, the fastest of the three. It pulled in a real-time market sentiment signal from X around the Phoenix multifamily market that neither Claude nor ChatGPT could (this is Grok's structural advantage). However, the prose required two passes of cleanup before it was IC ready: section headings drifted out of the requested format, and one risk paragraph contradicted itself between paragraphs three and four.
Test 2: Broker Fluff Detection
The same OM contained eight planted exaggerations: optimistic rent growth, a comp set with one outlier sale, a vague "value-add upside" claim with no specifics, an inflated occupancy assumption, and four others. We asked each model to flag every unsupported or optimistic claim.
Claude Opus 4.7: Identified 7 of 8 with specific quotes from the OM. This matches our prior finding in the AI underwriting speed test benchmark that Claude leads on document criticism.
ChatGPT GPT-5.5: Identified 5 of 8. Missed the comp set outlier and the vague upside claim.
Grok 4.3: Identified 4 of 8. Tended to summarize rather than challenge.
Test 3: Three-Way Underwriting Summary Table
We asked each model to produce a clean five-row underwriting table covering Purchase Price, NOI, Cap Rate, DSCR, and Cash on Cash, with footnotes explaining assumptions.
Claude Opus 4.7: Produced a clean HTML table with correct math and clearly labeled assumptions in the footnotes. NOI of $4.12 million on a $58.0 million purchase price gives a 7.10% cap rate (verified). DSCR calculated correctly at 1.32x using 70% LTV at a 6.5% rate.
ChatGPT GPT-5.5: Produced both a markdown table and a parallel Excel sheet. Math correct. Footnotes shorter than Claude's but accurate.
Grok 4.3: Produced a clean table. Math correct on cap rate and DSCR but used a slightly different operating expense assumption that was not flagged in the footnotes, requiring a follow-up to verify. For more detail on accuracy testing across models, see our Claude vs ChatGPT property valuation accuracy analysis.
Cost Per Memo at Volume
For a CRE shop producing thirty IC memos per month, average input around 80,000 tokens (OM plus T12 plus rent roll plus instructions) and average output around 6,000 tokens (three-page memo plus table), the monthly cost runs:
- Claude Opus 4.7: 80,000 x 30 = 2.4M input tokens at $5 per million = $12.00, plus 6,000 x 30 = 180,000 output tokens at $25 per million = $4.50. Total: $16.50 per month, or roughly $0.55 per memo.
- ChatGPT GPT-5.5: API pricing is comparable to GPT-5.4 at roughly $2 per million input and $10 per million output. Total: roughly $6.60 per month, or $0.22 per memo. (Per-seat ChatGPT Business at $25 per user per month is the more common pricing structure for shops.)
- Grok 4.3: 2.4M x $1.25 = $3.00 input, plus 180,000 x $2.50 = $0.45 output. Total: $3.45 per month, or roughly $0.115 per memo.
Token costs are rounding error compared to the partner-hour cost of rewriting a bad memo. The real economic comparison is about how often each model produces an IC ready first draft, and that math favors Claude Opus 4.7. For CRE professionals looking for hands-on guidance on building this kind of underwriting workflow, The AI Consulting Network specializes in exactly this implementation work.
Pricing Comparison for Volume CRE Shops
Independent market analysis from Artificial Analysis places Grok 4.3 on the cost-per-intelligence frontier in the second quarter of 2026, while Anthropic's Opus 4.7 announcement emphasizes the model's lead on long-horizon agentic tasks like multi-document underwriting. For a thirty-memo per month shop, the savings from running Grok for first drafts and using Claude only for IC polish can run 60 to 70% versus running every memo through Claude end to end.
Recommended Workflow for CRE Shops
Based on these results, here is the workflow we recommend to clients producing high IC memo volume:
- First draft: Grok 4.3 for speed and cost, especially when broker docs are long and the shop wants to test ten potential deals before committing partner time.
- Critical review pass: Claude Opus 4.7 for fluff detection, T12 versus pro forma reconciliation, and the actual IC memo prose. This is the highest-leverage use of Claude tokens.
- Spreadsheet build: ChatGPT GPT-5.5 with the Excel add-in to produce the supporting pro forma model in the same pass.
- Sentiment context: Grok 4.3 only when sponsor or market reputation matters (e.g., a new GP or a local market with recent insurance shocks).
This three-model stack is more expensive in API tokens than running everything through one model, but cheaper in partner hours and produces better IC ready output. CRE investors looking for a turnkey implementation can connect with The AI Consulting Network to build this workflow without the trial-and-error of testing each model independently.
Frequently Asked Questions
Q: Can Grok 4.3 replace Claude for IC memos entirely?
A: Not for shops that send memos to senior partners or external LPs. Grok 4.3 is the cost leader and writes acceptable first drafts, but its prose still requires cleanup, and its broker fluff detection trails Claude meaningfully. For internal screening memos at high volume, Grok alone is workable.
Q: Does the ChatGPT for Excel integration matter for IC memo workflow?
A: Yes, if your shop builds the supporting pro forma in Excel. ChatGPT GPT-5.5 with the Excel add-in produces the memo and the underlying model in a single pass, which neither Claude nor Grok can match natively. Sponsors who model in Google Sheets get the same benefit since the integration covers both.
Q: Which model is best for a solo sponsor doing fewer than five memos per month?
A: ChatGPT Plus or ChatGPT Business at $20 to $25 per month gives access to GPT-5.5 plus the Excel integration in a single subscription, which is the simplest path. Claude Pro at $20 per month is the alternative if memo writing quality matters more than spreadsheet building.
Q: Is Grok safe to use for confidential deal documents?
A: xAI maintains SOC 2 Type II auditing and enterprise data handling agreements that match Anthropic and OpenAI on the controls that matter for confidential deal documents. Sponsors should still review the data processing agreement before uploading sensitive documents, especially for institutional capital relationships.
Q: How often should we re-test these models for IC memo quality?
A: Every ninety days at minimum. Each of these three models has shifted measurably in the last sixty days alone (Claude Opus 4.7 launched April 16, Grok 4.3 launched April 17, GPT-5.5 became the ChatGPT default May 5). A workflow optimized in February 2026 is already out of date.