Real Estate Eval Benchmark
Every real estate process is a test: the same input, the same prompt and an answer key that was frozen before any system ran. The table shows each process against each system, including the processes where Реалист is behind.
Scored on November 2, 2026: 15 processes for Реалист, 39 for other systems. 39 of 39 processes have a frozen answer key (1,575 criteria), frozen on October 30, 2026. On the 15 processes where Реалист has a valid run, Реалист passes 66.9% of criteria (strict) and Claude 71.9%; Claude on all 33 of its cases passes 69.9%. Median time: Реалист 3,779 s, Claude 116 s. Both judge models are from one vendor (Anthropic); a judge from a second vendor and an expert grade are not yet available, so scores are preliminary.
Process by system
| Process | Client types | Реалист | Claude Opus 5.5 with web search | GPT | Gemini | Where Реалист is behind |
|---|---|---|---|---|---|---|
| P01 Deal search: mandate to shortlist | investor, broker, multi-asset | 81.0% of criteria (17 of 21), critical 6 of 7 | 81.0% of criteria (17 of 21), critical 6 of 7 | В ожидании | В ожидании | Not behind |
| P02 Deal scoring: filters and weights against a mandate | investor, multi-asset | No valid run: run did not finish, runner weekly limit | 50.0% of criteria (11 of 22), critical 3 of 6 | В ожидании | В ожидании | No valid Реалист run |
| P03 OM screening: red flags and arithmetic | investor, lender | 45.5% of criteria (10 of 22), critical 3 of 7 | 50.0% of criteria (11 of 22), critical 4 of 7 | В ожидании | В ожидании | Behind Claude Opus 5.5 with web search by 4.5 points: largest tenant |
| P04 Acquisition underwriting (real address, last trade, zoning, rates) | investor | 78.6% of criteria (22 of 28), critical 4 of 4 | 60.7% of criteria (17 of 28), critical 3 of 4 | В ожидании | В ожидании | Not behind |
| P05 Underwriting: T-12 vs tax record, NOI at price | investor, lender | No valid run: run did not finish, runner weekly limit | 80.0% of criteria (24 of 30), critical 19 of 21 | В ожидании | В ожидании | No valid Реалист run |
| P06 Rent roll audit vs leases | owner/AM, lender | No valid run: run did not finish, runner session limit | 73.9% of criteria (17 of 23), critical 8 of 8 | В ожидании | В ожидании | No valid Реалист run |
| P07 T-12 NOI reconciliation | owner/AM, lender | No valid run: run did not finish, runner session limit | 45.5% of criteria (10 of 22), critical 3 of 8 | В ожидании | В ожидании | No valid Реалист run |
| P08 Debt sizing: three constraints, DSCR, debt yield | lender, developer | No valid run: run did not finish, runner login expired | 80.0% of criteria (16 of 20), critical 8 of 9 | В ожидании | В ожидании | No valid Реалист run |
| P09 Refinance and defeasance | owner/AM, lender | 75.0% of criteria (15 of 20), critical 9 of 12 | 70.0% of criteria (14 of 20), critical 9 of 12 | В ожидании | В ожидании | Not behind |
| P10 Credit file: loan agreement abstract | lender | 72.4% of criteria (21 of 29), critical 7 of 9 | 82.8% of criteria (24 of 29), critical 8 of 9 | В ожидании | В ожидании | Behind Claude Opus 5.5 with web search by 10.3 points: floating-rate floor |
| P11 Credit file: leasehold / ground lease for lender | lender | No valid run: run did not finish, runner login expired | 67.7% of criteria (21 of 31), critical 5 of 7 | В ожидании | В ожидании | No valid Реалист run |
| P12 Data-room audit: deal book reconciled across documents | investor, lender | No valid run: run stopped by an engine configuration check | 52.9% of criteria (18 of 34), critical 5 of 7 | В ожидании | В ожидании | No valid Реалист run |
| P13 Lease abstraction (office) | owner/AM, broker | No valid run: run did not finish, runner login expired | 90.9% of criteria (20 of 22), critical 6 of 7 | В ожидании | В ожидании | No valid Реалист run |
| P14 Lease review for an occupier (sublease) | corporate | No valid run: run stopped by an engine configuration check | 91.7% of criteria (22 of 24), critical 8 of 8 | В ожидании | В ожидании | No valid Реалист run |
| P15 Lease comps for a tenant negotiation | corporate, broker | No valid run: run stopped by an engine configuration check | 65.0% of criteria (13 of 20), critical 7 of 11 | В ожидании | В ожидании | No valid Реалист run |
| P16 PSA review | investor, broker | 10.0% of criteria (2 of 20), critical 1 of 11 | 55.0% of criteria (11 of 20), critical 7 of 11 | В ожидании | В ожидании | Behind Claude Opus 5.5 with web search by 45.0 points: no analysis delivered, a clarifying question instead |
| P17 Title and environmental diligence | investor, lender | 80.0% of criteria (16 of 20), critical 7 of 9 | 75.0% of criteria (15 of 20), critical 6 of 9 | В ожидании | В ожидании | Not behind |
| P18 Sales comps (real address, kept and rejected) | investor, broker, lender | No valid run: run did not finish, runner login expired | 59.3% of criteria (16 of 27), critical 5 of 7 | В ожидании | В ожидании | No valid Реалист run |
| P19 Market rent and rent comps | owner/AM, broker | No valid run: run did not finish, runner login expired | 40.7% of criteria (11 of 27), critical 2 of 4 | В ожидании | В ожидании | No valid Реалист run |
| P20 Valuation: value opinion and entry price | investor, lender | 75.0% of criteria (21 of 28), critical 5 of 6 | 67.9% of criteria (19 of 28), critical 5 of 6 | В ожидании | В ожидании | Not behind |
| P21 BOV: broker opinion of value | broker, owner/AM | 28.6% of criteria (8 of 28), critical 1 of 3 | 60.7% of criteria (17 of 28), critical 2 of 3 | В ожидании | В ожидании | Behind Claude Opus 5.5 with web search by 32.1 points: value opinion above the answer band, numbers |
| P22 Feasibility: as-of-right buildable, units, parking | developer | 84.4% of criteria (27 of 32), critical 13 of 15 | 84.4% of criteria (27 of 32), critical 14 of 15 | В ожидании | В ожидании | Not behind |
| P23 Zoning and parking feasibility | developer | 90.0% of criteria (18 of 20), critical 14 of 15 | 85.0% of criteria (17 of 20), critical 14 of 15 | В ожидании | В ожидании | Not behind |
| P24 Development construction loan | developer, lender | 70.0% of criteria (14 of 20), critical 9 of 12 | 65.0% of criteria (13 of 20), critical 9 of 12 | В ожидании | В ожидании | Not behind |
| P25 Market study (supply, absorption, vacancy at delivery) | developer, investor, corporate | 60.0% of criteria (12 of 20), critical 6 of 11 | 70.0% of criteria (14 of 20), critical 9 of 11 | В ожидании | В ожидании | Behind Claude Opus 5.5 with web search by 10.0 points: market vintage mismatch, weakest submarket, conclusion |
| P26 Меморандум для инвесткомитета | investor, multi-asset | No valid run: run waited on a confirmation and the files it produced were not read | 60.0% of criteria (12 of 20), critical 4 of 8 | В ожидании | В ожидании | No valid Реалист run |
| P27 IC deck | investor, multi-asset | 65.0% of criteria (13 of 20), critical 7 of 11 | 85.0% of criteria (17 of 20), critical 9 of 11 | В ожидании | В ожидании | Behind Claude Opus 5.5 with web search by 20.0 points: deck format items, files not read |
| P28 Property passport: site, zoning, history | investor, broker, owner/AM, lender | No valid run: run not started | 93.3% of criteria (28 of 30), critical 7 of 7 | В ожидании | В ожидании | No valid Реалист run |
| P29 Property tax reassessment on sale | owner/AM, investor | 78.6% of criteria (22 of 28), critical 17 of 20 | 82.1% of criteria (23 of 28), critical 17 of 20 | В ожидании | В ожидании | Behind Claude Opus 5.5 with web search by 3.6 points: exclusions untested |
| P30 Rent-cap compliance and mark-to-market | owner/AM | No valid run: run not started, daily internal-run limit | 62.1% of criteria (18 of 29), critical 13 of 23 | В ожидании | В ожидании | No valid Реалист run |
| P31 Hotel management agreement (owner side) | owner/AM, lender | No valid run: run not started, daily internal-run limit | 63.3% of criteria (19 of 30), critical 8 of 9 | В ожидании | В ожидании | No valid Реалист run |
| P32 Multi-asset model operation: move exit cap on 20 assets | multi-asset | No valid run: run not started; Claude graded on the written answer only | 92.9% of criteria (13 of 14), critical 3 of 3 | В ожидании | В ожидании | No valid Реалист run |
| P33 Multi-asset model operation: fix a broken formula | multi-asset | No valid run: run not started; Claude graded on the written answer only | 76.5% of criteria (13 of 17), critical 2 of 3 | В ожидании | В ожидании | No valid Реалист run |
| P34 Property data accuracy: 10 recorded figures per address (DATABENCH) | all | В ожидании | 76.7% of figures (468 of 610) | В ожидании | В ожидании | В ожидании |
| P35 Owner and people search (owner of record, controlling person, contact path) | broker, investor | No valid run: run not started, daily internal-run limit | 47.4% of criteria (18 of 38) | В ожидании | В ожидании | No valid Реалист run |
| P36 Signals: distress and market events on a watched set | investor, lender | No valid run: run not started, daily internal-run limit | 59.5% of criteria (22 of 37) | В ожидании | В ожидании | No valid Реалист run |
| P37 Portfolio monitoring: what changed since last review | owner/AM, multi-asset, lender | No valid run: run not started, daily internal-run limit | 88.0% of criteria (22 of 25) | В ожидании | В ожидании | No valid Реалист run |
| P38 OM and BOV deck authoring from source documents | broker | No valid run: run not started, daily internal-run limit | 80.8% of criteria (21 of 26) | В ожидании | В ожидании | No valid Реалист run |
| P39 Deal search from listings against a buy box (live inventory) | investor, broker | No valid run: run not started, daily internal-run limit | 84.6% of criteria (22 of 26) | В ожидании | В ожидании | No valid Реалист run |
All 39 processes have a frozen case and answer key. GPT and Gemini columns stay pending until those models are run under the same rules.
Where Реалист is behind
A process lands here when Реалист scores below the best other system on the share of criteria passed. The failed criteria class is shown with it. Nothing is removed from this list to make the table look better.
P16 PSA reviewBehind Claude Opus 5.5 with web search by 45.0 points: no analysis delivered, a clarifying question instead
P21 BOV: broker opinion of valueBehind Claude Opus 5.5 with web search by 32.1 points: value opinion above the answer band, numbers
P27 IC deckBehind Claude Opus 5.5 with web search by 20.0 points: deck format items, files not read
P10 Credit file: loan agreement abstractBehind Claude Opus 5.5 with web search by 10.3 points: floating-rate floor
P25 Market study (supply, absorption, vacancy at delivery)Behind Claude Opus 5.5 with web search by 10.0 points: market vintage mismatch, weakest submarket, conclusion
P03 OM screening: red flags and arithmeticBehind Claude Opus 5.5 with web search by 4.5 points: largest tenant
P29 Property tax reassessment on saleBehind Claude Opus 5.5 with web search by 3.6 points: exclusions untested
Time: Реалист is slower on every matched case, median 3,779 s against 116 s for Claude.
Where Реалист is ahead or level
P04 Acquisition underwriting (real address, last trade, zoning, rates)Ahead by 17.9 points
P20 Valuation: value opinion and entry priceAhead by 7.1 points
P09 Refinance and defeasanceAhead by 5.0 points
P17 Title and environmental diligenceAhead by 5.0 points
P23 Zoning and parking feasibilityAhead by 5.0 points
P24 Development construction loanAhead by 5.0 points
P01 Deal search: mandate to shortlistУровень
P22 Feasibility: as-of-right buildable, units, parkingУровень
Метод
- One case per process and client type. The list has 39 processes for 7 client types: investor, lender, developer, owner and asset manager, broker, corporate occupier and multi-asset holder.
- The case file holds the prompt, the documents, the answer key and the tolerance for every number. It is frozen with a hash of each file before the first run. A run counts only if the file it used has the frozen hash. A key is never edited after the first run: a correction is a new version and the old results stay.
- Every system gets the same input and the same prompt, once, with no hints and no retries. Реалист runs on its own engine. The general models run with web search in an empty working directory, with no memory and no added skills.
- Numbers are checked against the key with the case tolerance (exact, absolute or percent). This step is deterministic and uses no model.
- Every other criterion is graded by two judge models. The judges are not told which system wrote the answer, and a criterion counts as passed only when both pass it. Both judge models are from one vendor (Anthropic); a judge from a second vendor and an expert grade are not yet available, so scores are preliminary. Where a human expert has graded the same sheet, the expert’s grade will win over the judges.
- Each case reports the share of criteria passed, whether every critical criterion passed, number accuracy and run time. The table is process by system.
- The team that wrote a key does not score it.
Answer keys frozen October 30, 2026.
Prompts
The prompt is published next to the result. Where the documents of a case may not be shown, the page shows the score and the class of the criteria that failed, not the documents. Typography is normalized for display (long dashes shown as commas); the hash covers the original file.
P06 Rent roll audit vs leasesTake the rent roll and the lease folder for this Austin property and tell me: occupancy, in-place average rent per unit and per SF, loss-to-lease against the rents the leases actually carry, how many units expire in the next 6 and 12 months, and every unit where the roll and the leases disagree. Flag anything that would change what I underwrite.
P12 Data-room audit: deal book reconciled across documentsHere's the full book on Bayou Crest Apartments in Houston, OM, rent roll, T-12, sample leases, the management agreement. Before this goes to IC I need the documents reconciled against each other: everywhere the OM's story doesn't match the roll, the T-12, the leases, what each gap is worth, and the questions I should send the broker. Then screen it against our buy box and tell me honestly how many criteria it passes.
P22 Feasibility: as-of-right buildable, units, parkingRun the feasibility on 1412 Quillfeather Place. Tell me the as-of-right residential buildable area, which control binds it and why, the unit count at a 750 SF average, whether any parking is required, what the inclusionary option is worth and what it costs, and whether the broker's 19,300 SF and the community facility number are right. Show the arithmetic.
P32 Multi-asset model operation: move exit cap on 20 assetsMove the exit cap rate out 50 bps on the first 20 assets.
P33 Multi-asset model operation: fix a broken formulaSomeone pasted a value over the NOI formula in one month of the second asset. Put the formula back.
The prompts of the other cases are published with their first scored result.