Skip to content
HomieBench v3Kimi K3 · Directional July 2026 update

HomieBench v3: the best AI for realtors, ranked by finished work.

HomieBench compares the models behind today’s best realtor AI assistants across offers, CRM, lead generation, marketing, CMAs, showings, MLS research, document review, and closings. Kimi K3 takes the projected v3 lead—but it does not win every kind of realtor work.

Homies ResearchEditorial forecast · ±3-point uncertainty
The v3 signal
Kimi leads. Routing still matters.
1
Best projected overall
Kimi K3
96.6
2
Browser-heavy realtor work
Kimi K3
98.0
3
Best modeled cost / benefit
Kimi K3
$0.153

Kimi K3 is newly available. Its v3 placement is a scenario projection from official model and pricing evidence plus HomieBench priors—not a claim that the repeatable identical-harness run is already complete.

frontier models
8
realtor workflows
100
job families
8
shared harness
1
Share Post LinkedIn Email
92%

of 225 surveyed U.S. NAR-member agents use AI now or plan to.

Adoption is no longer the hard question. Trust is: 63% named output accuracy as a top concern and 49% named compliance or legal issues in the same n=225 survey. HomieBench is built around completed, reviewable jobs—not impressive chat demos. NAR / RPR survey

Results at a glance

Best AI models for real estate agents in 2026

Quality, reliability, browser use, and completed-job economics are different questions. These v3 scores are scenario projections calibrated to one shared realtor task map and clearly separated from completed identical-harness runs.

Projected overall leader

Kimi K3

The v3 leader across the full realtor workflow mix, with a 96.6 projected score.

Browser + tool leader

Kimi K3

Projected first for showing booking, CRM management, MLS work, and offer assembly.

Best cost / benefit

Kimi K3

Projected first after balancing direct AI spend, finished-work quality, and human review burden.

Judgment + content leader

Claude Fable 5

Still projected ahead on marketing quality and nuanced high-stakes realtor judgment.

Interactive leaderboard

Change the job. Watch the ranking change.

All scores / 100 · Scenario projection
RankModelScore
  1. 1
    Kimi K3 New
    Moonshot AIScenario projection
    96.6
    out of 100
  2. 2
    Claude Fable 5
    AnthropicScenario projection
    95.5
    out of 100
  3. 3
    GPT-5.6 Sol
    OpenAIScenario projection
    95.4
    out of 100
  4. 4
    Claude Sonnet 5
    AnthropicScenario projection
    94.3
    out of 100
  5. 5
    Muse Spark 1.1
    MetaScenario projection
    91.1
    out of 100
  6. 6
    Gemini 3.5 Flash
    GoogleScenario projection
    90.9
    out of 100
  7. 7
    Grok 4.5
    SpaceXAIScenario projection
    90.0
    out of 100
  8. 8
    GLM-5.2
    Z.aiScenario projection
    85.6
    out of 100

Changing the job changes the order. That is the point: the best model for offer strategy is not automatically the best model for showing coordination, content, or cost.

All HomieBench scenario-projection scores by model and job family
ModelOverallLead generation & prospectingCRM & client communicationsMarketing & contentShowings & coordinationProperty, market & document intelligenceOffers & negotiationTransactions, closings & complianceBack office & client deliverables
Kimi K396.69598949997989497
Claude Fable 595.59494969496979696
GPT-5.6 Sol95.49496939795959795
Claude Sonnet 594.39294959594949595
Muse Spark 1.191.19092919491899093
Gemini 3.5 Flash90.98991919293899092
Grok 4.590.09091909389888893
GLM-5.285.68586858887848388
Best AI by realtor job

Which AI is best for each real estate workflow?

A strong realtor AI assistant should route the job instead of asking one model to be the best researcher, copywriter, coordinator, analyst, and negotiator at once.

Browser + multi-tool work

Kimi K3

Showing booking, CRM management, MLS and property research, offer assembly, and long multi-step tool execution.

High-stakes judgment

Claude Fable 5

Nuanced negotiation strategy, polished client work, marketing judgment, and decisions where subtlety matters.

Closings + compliance

GPT-5.6 Sol

Reliable transaction follow-through, deadline tracking, structured outputs, and compliance-sensitive workflows.

Kimi K3 vs. Claude Fable 5 for real estate

Why Kimi K3 leads realtor browser tasks in HomieBench v3

Kimi K3’s official 1M-token context, native visual understanding, structured output, dynamic tool loading, and tool-choice controls create a strong prior for long browser-and-system workflows. That matters when an AI must move from MLS research to a CRM record, showing request, document, or offer package without losing state.

The forecast is deliberately not a sweep. Fable 5 remains ahead in marketing and subtle client-facing judgment, while GPT-5.6 Sol remains ahead in closings and compliance. Kimi wins the weighted overall mix because realtors do a lot of tool-mediated work.

Kimi K3 compared with Claude Fable 5 on browser-heavy realtor task categories
Realtor browser taskKimi K3Fable 5Kimi lead
Showing booking & coordination9994+5
CRM management & communications9894+4
MLS, property & document work9796+1
Offer writing & assembly9897+1
All categories at once

How every model ranks across the eight realtor job families

Each bubble is one model’s projected category score. For priced models, area adds a model-level provider list-price estimate.

64 model-category scores
Axis zoomed to 80–100

Higher dots mean higher projected quality. Each model keeps the same bubble size in every category: area shows its average single-pass provider list-price estimate across the six sample jobs—not that category’s task cost, tools, retries, subscriptions, or Homies effective cost. GLM-5.2 is outlined and not size-scaled because self-hosting cost varies. Hover or focus a dot to see its model and values.

  • Kimi K3
  • Fable 5
  • GPT-5.6
  • Sonnet 5
  • Muse 1.1
  • Gemini 3.5
  • Grok 4.5
  • GLM-5.2
Average provider list-price estimatelower → higher· outlined = variable

Swipe to explore all eight categories →

HomieBench projected model quality by realtor category with average provider list-price estimate by modelEight categories appear on the horizontal axis and scenario-projection score from 80 to 100 appears on the vertical axis. Each colored bubble is one model. Bubble area uses that model’s average single-pass provider list-price estimate across six sample jobs and remains the same across categories. GLM-5.2 has variable self-hosted cost and is not size-scaled. Use arrow keys to move between bubbles and inspect model, category, score, and cost.

Bubble area uses each model’s average provider list-price estimate across the six published sample jobs. It excludes tools, retries, subscription allocation, and human rescue; it is not observed completed-job cost.

HomieBench projected category scores and average provider list-price estimates across the six published sample jobs
CategoryModelProjected scoreAverage single-pass provider list-price estimate across six sample jobs
Lead generation & prospectingKimi K395$1.17
Lead generation & prospectingClaude Fable 594$3.91
Lead generation & prospectingGPT-5.6 Sol94$2.21
Lead generation & prospectingClaude Sonnet 592$0.782
Lead generation & prospectingMuse Spark 1.190$0.388
Lead generation & prospectingGemini 3.5 Flash89$0.662
Lead generation & prospectingGrok 4.590$0.580
Lead generation & prospectingGLM-5.285Variable self-hosted cost
CRM & client communicationsKimi K398$1.17
CRM & client communicationsClaude Fable 594$3.91
CRM & client communicationsGPT-5.6 Sol96$2.21
CRM & client communicationsClaude Sonnet 594$0.782
CRM & client communicationsMuse Spark 1.192$0.388
CRM & client communicationsGemini 3.5 Flash91$0.662
CRM & client communicationsGrok 4.591$0.580
CRM & client communicationsGLM-5.286Variable self-hosted cost
Marketing & contentKimi K394$1.17
Marketing & contentClaude Fable 596$3.91
Marketing & contentGPT-5.6 Sol93$2.21
Marketing & contentClaude Sonnet 595$0.782
Marketing & contentMuse Spark 1.191$0.388
Marketing & contentGemini 3.5 Flash91$0.662
Marketing & contentGrok 4.590$0.580
Marketing & contentGLM-5.285Variable self-hosted cost
Showings & coordinationKimi K399$1.17
Showings & coordinationClaude Fable 594$3.91
Showings & coordinationGPT-5.6 Sol97$2.21
Showings & coordinationClaude Sonnet 595$0.782
Showings & coordinationMuse Spark 1.194$0.388
Showings & coordinationGemini 3.5 Flash92$0.662
Showings & coordinationGrok 4.593$0.580
Showings & coordinationGLM-5.288Variable self-hosted cost
Property, market & document intelligenceKimi K397$1.17
Property, market & document intelligenceClaude Fable 596$3.91
Property, market & document intelligenceGPT-5.6 Sol95$2.21
Property, market & document intelligenceClaude Sonnet 594$0.782
Property, market & document intelligenceMuse Spark 1.191$0.388
Property, market & document intelligenceGemini 3.5 Flash93$0.662
Property, market & document intelligenceGrok 4.589$0.580
Property, market & document intelligenceGLM-5.287Variable self-hosted cost
Offers & negotiationKimi K398$1.17
Offers & negotiationClaude Fable 597$3.91
Offers & negotiationGPT-5.6 Sol95$2.21
Offers & negotiationClaude Sonnet 594$0.782
Offers & negotiationMuse Spark 1.189$0.388
Offers & negotiationGemini 3.5 Flash89$0.662
Offers & negotiationGrok 4.588$0.580
Offers & negotiationGLM-5.284Variable self-hosted cost
Transactions, closings & complianceKimi K394$1.17
Transactions, closings & complianceClaude Fable 596$3.91
Transactions, closings & complianceGPT-5.6 Sol97$2.21
Transactions, closings & complianceClaude Sonnet 595$0.782
Transactions, closings & complianceMuse Spark 1.190$0.388
Transactions, closings & complianceGemini 3.5 Flash90$0.662
Transactions, closings & complianceGrok 4.588$0.580
Transactions, closings & complianceGLM-5.283Variable self-hosted cost
Back office & client deliverablesKimi K397$1.17
Back office & client deliverablesClaude Fable 596$3.91
Back office & client deliverablesGPT-5.6 Sol95$2.21
Back office & client deliverablesClaude Sonnet 595$0.782
Back office & client deliverablesMuse Spark 1.193$0.388
Back office & client deliverablesGemini 3.5 Flash92$0.662
Back office & client deliverablesGrok 4.593$0.580
Back office & client deliverablesGLM-5.288Variable self-hosted cost
LeadBench

Lead generation & prospecting

Finding, prioritizing, qualifying, nurturing, and booking the right buyer and seller opportunities.

Sample task

Build a compliant 30-day reactivation campaign for 200 past clients and cold leads, prioritize the call list, and create CRM tasks.

Per-model score
Kimi K3Best
95
Claude Fable 5
94
GPT-5.6 Sol
94
Claude Sonnet 5
92
Muse Spark 1.1
90
Grok 4.5
90
Gemini 3.5 Flash
89
GLM-5.2
85
How we test the models

The model is the brain. The Homies harness is the toolbox.

Models do the reasoning. Homies gives them the real-estate context, tools, memory, permissions, and review gates needed to finish the job. Holding that toolbox constant is the only useful way to compare the engines.

Homies AI harness

One realtor toolbox. Any reasoning engine.

The HomieBench design gives every model the same real estate context, tools, and safeguards, then measures the finished work it can produce.

01 · Swappable model

The reasoning engine

The model interprets the request, reasons through the case, and decides which tool to use next.

  • Kimi K3
  • Claude Fable 5
  • GPT-5.6 Sol
  • Claude Sonnet 5
  • Muse Spark 1.1
  • Gemini 3.5 Flash
  • Grok 4.5
  • GLM-5.2
Better models can swap in without rebuilding the agent’s workflow.
02 · Homies AI
Same harness every run

The realtor toolbox

The harness supplies the data access, integrations, memory, and operating rules that turn a smart answer into completed real estate work.

Tools each model can use
  • CRM
  • Property data
  • Email
  • Calendar
  • Documents
  • Research
  • Calculator
  • Publishing
Always attached
  • Real estate context
  • Durable memory
  • Permissions
  • Review gates
03 · Agent-ready results

Finished work, not chat

Human review gate

The agent approves advice, commitments, and anything client-facing before it goes out.

  • CRM updated and follow-up drafted
  • Showing tour booked and confirmed
  • CMA and listing presentation ready
  • Offer written, summarized, and flagged
  • Property campaign packaged for approval
The model is the reasoning engine. The Homies harness is the toolbox: context, memory, permissions, integrations, and human review gates. Our mission is to build the best AI toolbox for real estate agents and keep testing every model that could make it better.
Why the harness matters

A brilliant model without the right property data, forms, CRM context, tools, and authority limits can still produce unusable work. A strong harness makes the work grounded, repeatable, reviewable, and connected to the agent’s actual business.

Read the full realtor harness guide

Our companion research paper defines the nine components of a production AI harness and the four operating levels a realtor can run it at. Most individual agents belong at Level 1 or 2, where consequential work still clears human approval.

What is an AI harness? The realtor’s guide
Cost per completed realtor outcome

The cheapest tokens are rarely the cheapest offer, showing, or CMA

HomieBench v3 estimates direct AI and incremental tool spend per successful production-sized realtor task. Human verification remains essential, but its labour value is disclosed separately instead of being hidden inside the cents-per-task figure.

Realtor unit economics

Estimated AI cost per offer, showing, CMA, MLS brief, and CRM follow-up

These headline numbers estimate direct model and incremental tool spend for a successful production-sized run, including retry risk. Human review time is shown separately and is not added to the cents-per-task figure.

Projected overall leader
Kimi K3
96.6 / 100 across 8 job families
Browser + tool score
98.0
CRM, showings, MLS/property work, and offer assembly
Average direct AI cost
$0.15
Model + incremental tools across the five representative outcomes below
HomieBench v3 estimated direct AI and incremental tool cost per successful realtor outcome by model
Rank and modelOverallValue index$ / CRM follow-up$ / Showing booked$ / MLS brief$ / Offer written$ / CMA completedAverage
1Kimi K3Best valueMoonshot AI
96.6100$0.04$0.08$0.14$0.18$0.33$0.15
2GPT-5.6 SolOpenAI
95.478$0.07$0.12$0.23$0.32$0.56$0.26
3Claude Fable 5Anthropic
95.574$0.12$0.21$0.40$0.55$0.99$0.45
4Claude Sonnet 5Anthropic
94.369$0.03$0.06$0.11$0.14$0.24$0.12
5Gemini 3.5 FlashGoogle
90.953$0.03$0.05$0.09$0.13$0.21$0.10
6Muse Spark 1.1Meta
91.152$0.02$0.04$0.07$0.09$0.16$0.08
7Grok 4.5SpaceXAI
90.046$0.03$0.05$0.10$0.12$0.22$0.10
GLM-5.2Z.ai
85.6VariableVariableVariableVariableVariableVariableVariable

Headline formula: (published model token cost + incremental per-run tool charges) ÷ projected successful completion. Production-sized prompts are used here, not the much larger benchmark case files. The value index separately combines projected overall quality with the fully loaded cost of review time valued at $60/hour, then normalizes the leader to 100.

Human verification remains necessary. It is modeled separately at approximately 1–7 minutes for Kimi K3 across these tasks and valued internally at $60/hour, but is not included in the displayed AI cost. Also excludes fixed CRM, MLS, showing-platform, forms, signatures, subscriptions, taxes, and enterprise support.

The value frontier

Projected quality vs. loaded cost per completed outcome

HomieBench v3 projected overall score against average loaded cost per completed realtor outcome, by modelScatter chart. The horizontal axis is average loaded cost per completed outcome in dollars, combining direct AI spend with review time valued at $60 per hour. The vertical axis is projected overall score out of 100. Kimi K3 holds the best position at 96.6 points and $3.71 per outcome. A data table with every value follows the chart.
HomieBench v3 projected overall score, average loaded cost per completed outcome, and value index by model
ModelProjected overall score out of 100Average loaded cost per completed outcomeValue index (leader = 100)
Kimi K396.6$3.71100
GPT-5.6 Sol95.4$4.7178
Claude Fable 595.5$4.9474
Claude Sonnet 594.3$5.2469
Gemini 3.5 Flash90.9$6.6453
Muse Spark 1.191.1$6.7552
Grok 4.590.0$7.5846
Loaded cost adds modeled human review time, valued at $60/hour, to direct model and incremental tool spend per successful outcome. Kimi K3 sits alone on the frontier with the highest projected score and the lowest loaded cost of the 7 priced models. GLM-5.2 is excluded because self-hosted economics vary. Scenario projections, not completed identical-harness runs.

Human review burden

Average human review minutes per completed outcome

Kimi K3
3.6 min
GPT-5.6 Sol
4.5 min
Claude Fable 5
4.5 min
Claude Sonnet 5
5.1 min
Gemini 3.5 Flash
6.5 min
Muse Spark 1.1
6.7 min
Grok 4.5
7.5 min
Modeled minutes an agent spends verifying each completed outcome, averaged across the five representative outcomes and valued at $60/hour in the loaded costs above. Review time, not tokens, dominates loaded cost, which is why the cheapest tokens rarely produce the cheapest completed job. GLM-5.2 is excluded because self-hosted economics vary.
Access-route economics

Provider list price and Homies effective cost are still separate questions

List price is one input; the access route decides the rest. The explorer below keeps the two views separate, and the complete real estate AI integration stack maps which realtor systems those per-run tool calls actually reach.

Qualitative Homies cost signal

A directional grade—not a measured dollar result—based on connected-plan allocation, limits, retries, and tool fees.

  1. #1
    Kimi K3

    Projected first-pass completion and browser/tool efficiency offset its mid-tier token price.

    A+
    qualitative grade
  2. #2
    GPT-5.6 Sol

    Strong connected-account economics and the projected reliability leader for closing workflows.

    A
    qualitative grade
  3. #3
    Gemini 3.5 Flash

    Strong price-performance prior for high-volume multimodal support work.

    A
    qualitative grade
  4. #4
    Muse Spark 1.1

    Strong agentic prior with lower reported pricing than Grok 4.5.

    A
    qualitative grade
  5. #5
    Grok 4.5

    Efficient agentic route, but pricier than Muse on the reported launch card.

    B+
    qualitative grade
  6. #6
    Claude Sonnet 5

    Strong quality-to-cost balance, subject to post-intro pricing.

    B
    qualitative grade
  7. #7
    Claude Fable 5

    Premium route reserved for high-consequence judgment.

    D
    qualitative grade
  8. GLM-5.2

    Infrastructure, utilization, support, and hosting choices determine cost.

    Variable
    qualitative grade

How Homies frames effective cost

Framework—not run data

(allocated plan cost + metered overage + tool fees + retry spend) ÷ successful jobs

Kimi K3 ranks first in this qualitative v3 estimate because its projected first-pass completion and browser/tool efficiency offset its mid-tier token price. GPT-5.6 remains a strong connected-account route. OAuth authenticates a connection; it does not make inference free. Plan fees, limits, overages, and separate API billing still apply.

Provider prices updated July 16, 2026. Kimi K3 uses Moonshot AI’s published $3 input / $15 output per million-token cache-miss pricing. Muse Spark 1.1 pricing is launch reporting pending confirmation in the Meta console. Claude Sonnet 5 uses its introductory rate. Taxes, regional pricing, search/tool charges, and subscription fees are excluded from token-only examples.

The realtor task suite

What is inside all 100 HomieBench workflows?

The headline categories stay simple. Underneath them is the real work of running a real estate business—from the first lead to years after closing, with the files, trades, deadlines, and client judgment in between.

0113 workflows

Lead generation & prospecting

Finding, prioritizing, qualifying, nurturing, and booking the right buyer and seller opportunities.

Projected leader Kimi K3 · 95
0213 workflows

CRM & client communications

Keeping the database clean, the pipeline current, and every client conversation accurate and useful.

Projected leader Kimi K3 · 98
0313 workflows

Marketing & content

Creating accurate listing marketing and on-brand social, email, video, advertising, and nurture content.

Projected leader Fable 5 · 96
0412 workflows

Showings & coordination

Scheduling tours, optimizing routes, coordinating listing offices and trades, and keeping every party informed.

Projected leader Kimi K3 · 99
0513 workflows

Property, market & document intelligence

Researching properties and markets, building valuations, and reviewing inspections, title, zoning, HOA, and condo records.

Projected leader Kimi K3 · 97
0612 workflows

Offers & negotiation

Structuring, drafting, explaining, comparing, presenting, countering, and negotiating offers under agent approval.

Projected leader Kimi K3 · 98
0712 workflows

Transactions, closings & compliance

Managing conditions, escrow/deposits, lenders, lawyers, title, insurance, walkthroughs, privacy, and compliance.

Projected leader GPT-5.6 · 97
0812 workflows

Back office & client deliverables

Organizing files, extracting documents, preparing signatures, coordinating collaborators, and building reports and presentations.

Projected leader Kimi K3 · 97
Complete coverage matrix

Every workflow in the HomieBench map

Open a family to inspect every task and its review level. High-stakes work must clear human approval and all critical criteria.

01Lead generation & prospecting13 workflows · 10% of quality score
WorkflowWhat the agent must finishReview level
Lead-source importImport leads from portals, ads, open houses, referrals, and spreadsheets without losing source data.Routine
Ideal-client profileDefine the audience, geography, property type, motivation, and qualification signals for a campaign.Review required
Farm-area prospect listBuild a prioritized geographic farm list with reasons and a compliant next action.Review required
Seller-intent signalsIdentify contacts showing plausible move, equity, life-event, or engagement signals without inventing facts.Review required
Buyer-intent signalsPrioritize buyers by activity, timeframe, financing readiness, and property fit.Review required
Expired-listing outreachResearch an expired listing and prepare a compliant, personalized multi-touch approach.Review required
FSBO outreachPrepare respectful owner outreach, value framing, discovery questions, and follow-up timing.Review required
Past-client reactivationFind dormant relationships and create a useful re-engagement reason such as an equity or CMA update.Routine
Database nurture segmentsGroup contacts by relationship, intent, timing, market, and next-best campaign.Routine
Outbound call list and scriptPrioritize a daily call list and draft context-aware openings, questions, and voicemail.Review required
Multi-channel prospecting sequenceCreate coordinated email, SMS, call, and social touches with timing and stop rules.Review required
Lead qualification and scoreAssess motivation, agency status, timeframe, financing, fit, and follow-up urgency.Review required
Appointment setting and handoffOffer suitable times, book the meeting, create the CRM event, and prepare the agent brief.Review required
02CRM & client communications13 workflows · 15% of quality score
WorkflowWhat the agent must finishReview level
Contact deduplicationMerge duplicate people and households while preserving attribution, notes, consent, and history.Routine
Contact enrichmentStructure known preferences, relationships, properties, and communication details without guessing.Review required
Conversation summarizationTurn calls, emails, and messages into a factual timeline, decisions, concerns, and next actions.Routine
Lifecycle and stage classificationPlace contacts and opportunities in the correct stage using explicit evidence.Routine
Next-best actionRecommend the most useful next step, owner, channel, and due date for each relationship.Review required
Inbound inquiry responseDraft a fast, helpful reply that answers known facts, asks useful questions, and avoids commitments.Review required
Buyer discovery briefCapture needs, trade-offs, financing, timing, decision-makers, and search boundaries.Review required
Seller discovery briefCapture motivation, property context, timing, condition, expectations, and decision criteria.Review required
Client email draftingWrite clear, accurate, on-brand email from CRM and transaction context.Review required
Client SMS draftingWrite concise, context-aware text messages with correct tone and no invented promises.Review required
Long-term nurture planCreate relationship-first follow-up that stays useful across a long buying or selling horizon.Review required
Objection responsePrepare calibrated responses to fee, timing, pricing, competition, and process objections.High stakes
Pipeline health reportSummarize conversion risk, stalled opportunities, overdue work, and coaching priorities.Routine
03Marketing & content13 workflows · 10% of quality score
WorkflowWhat the agent must finishReview level
MLS remarksDraft accurate, compliant public remarks from verified property facts and approved positioning.Review required
Property highlight sheetTurn features, improvements, rooms, and lifestyle context into a scannable fact sheet.Review required
Listing brochureBuild structured brochure copy, hierarchy, calls to action, and proof points in the agent brand.Review required
Listing landing pageCreate the page outline, property story, feature modules, lead capture, and SEO copy.Review required
Property email campaignDraft announcement, open-house, update, and follow-up emails for the right audience.Review required
Social content calendarPlan useful listing, market, education, community, and personal-brand posts.Routine
Social captionsCreate platform-aware captions, hooks, calls to action, and compliant hashtags.Review required
Carousel creationTurn a market insight or property story into a clear slide-by-slide social carousel.Review required
Reel and video scriptWrite short-form and long-form real estate video scripts with shots, hooks, and captions.Review required
Paid-ad campaignBuild audience, creative angle, copy variants, landing-page match, and measurement plan.High stakes
Brand-voice rewriteAdapt content to the agent's approved tone without changing facts or compliance meaning.Routine
Newsletter productionAssemble market, listing, client, and community content into a useful recurring newsletter.Review required
Performance repurposingAnalyze approved content performance and turn strong ideas into new channel-native assets.Routine
04Showings & coordination12 workflows · 10% of quality score
WorkflowWhat the agent must finishReview level
Showing availability planReconcile client, agent, property, notice, occupancy, and travel constraints.Review required
Listing-office showing requestPrepare or place the authorized request with correct party, property, time, and conditions.Review required
Tour route optimizationOrder properties for drive time, appointment windows, breaks, and client priorities.Routine
Tour confirmation packageSend the itinerary, access notes, property links, timing, and preparation reminders.Review required
Showing rescheduleResolve conflicts, re-contact parties, update calendars, and preserve the rest of the route.Review required
Showing feedbackCollect, summarize, and route useful buyer feedback without exposing confidential information.Review required
Open-house operationsPrepare schedule, signage, registration, safety, follow-up, and seller reporting.Review required
Calendar blockingCreate accurate appointments, buffers, travel time, reminders, and linked records.Routine
Vendor bookingCoordinate approved photographers, stagers, cleaners, contractors, and measurements.Review required
Inspection coordinationBook the inspector, align parties, share access instructions, and track the report.High stakes
Appraisal accessCoordinate appraisal timing, property access, contacts, and the approved information package.Review required
Client tour briefPrepare a concise mobile itinerary with property fit, verified facts, questions, and flags.Review required
05Property, market & document intelligence13 workflows · 15% of quality score
WorkflowWhat the agent must finishReview level
Listing fact verificationExtract and reconcile facts across the listing, tax record, disclosures, and source documents.High stakes
Neighbourhood researchPrepare sourced context on amenities, mobility, housing, plans, and client-relevant trade-offs.Review required
Market trend reportAnalyze inventory, absorption, pricing, days on market, and segment-level movement.Review required
Comparable selectionChoose defensible sold, active, expired, and leased comps with inclusion reasons.High stakes
Comparable adjustmentsAdjust for time, size, condition, lot, parking, features, and location with uncertainty.High stakes
CMA productionBuild the evidence table, pricing range, positioning narrative, and seller-ready report.High stakes
Home evaluationEstimate a value range, confidence, key drivers, missing data, and next validation steps.High stakes
Investor analysisModel revenue, expenses, financing, cash flow, cap rate, sensitivity, and risks.High stakes
Rental estimateSelect rental evidence, adjust for features and timing, and explain a supportable range.High stakes
Zoning and permit researchFind applicable zoning, permits, constraints, and questions for the right authority.High stakes
Tax, title, and survey reviewSummarize source documents, inconsistencies, easements, boundaries, and referral questions.High stakes
HOA and condo document reviewReview bylaws, minutes, budgets, reserves, fees, insurance, restrictions, and litigation signals.High stakes
Home-inspection reviewSummarize findings by urgency, cost uncertainty, specialist need, and negotiation relevance.High stakes
06Offers & negotiation12 workflows · 15% of quality score
WorkflowWhat the agent must finishReview level
Buyer offer strategyTranslate goals, competition, financing, risk tolerance, and property facts into a strategy.High stakes
Term recommendationRecommend price, deposit, closing, inclusions, conditions, and expiry with trade-offs.High stakes
Clause selectionChoose only broker-approved clauses that match instructions, jurisdiction, and deal facts.High stakes
Deadline calculationCalculate expiry, condition, deposit, notice, and closing dates with calendar rules.High stakes
Deposit and closing planCheck deposit mechanics, funding timing, closing feasibility, and client explanation.High stakes
Offer draftingPrepare a review-ready purchase or lease offer from verified instructions and approved forms.High stakes
Buyer offer explanationExplain terms, obligations, risks, alternatives, and approval points in plain language.High stakes
Seller offer summaryPresent price, terms, conditions, timing, risks, and net implications without hiding trade-offs.High stakes
Multiple-offer comparisonNormalize offers side by side and flag material differences, gaps, and decision points.High stakes
Counteroffer draftingPrepare the approved changes, rationale, timing, and client/counterparty communication.High stakes
Negotiation planMap priorities, leverage, concessions, signals, limits, and fallback paths for agent approval.High stakes
Amendment and waiver supportTrack the requested change, authority, form, dates, dependencies, and signature status.High stakes
07Transactions, closings & compliance12 workflows · 15% of quality score
WorkflowWhat the agent must finishReview level
Accepted-offer extractionTurn the executed agreement into parties, dates, conditions, obligations, and a deal ledger.High stakes
Condition and contingency trackerTrack each requirement, responsible party, evidence, deadline, and escalation path.High stakes
Escrow and deposit trackingMonitor instructions, receipt, trust/escrow status, deadlines, and exceptions without moving funds.High stakes
Financing coordinationPrepare the lender package, track approval steps, surface gaps, and keep parties aligned.High stakes
Appraisal issue briefSummarize a valuation gap, contract implications, options, and questions for licensed advisers.High stakes
Inspection resolutionConvert findings into specialist referrals, repair/credit options, deadlines, and client decisions.High stakes
Lawyer, title, and insurance liaisonShare the approved package, track questions, and route issues to the right professional.High stakes
Closing timelineCreate a dated plan for financing, legal, insurance, utilities, movers, walkthrough, and keys.High stakes
Final walkthroughPrepare the checklist, evidence capture, deficiency routing, and urgent escalation plan.High stakes
Key handover and post-closeCoordinate possession, keys, closing communication, record updates, and relationship follow-up.Review required
Fair-housing and advertising reviewFlag protected-class targeting, steering, exclusionary language, and risky claims.High stakes
Privacy and record-retention reviewMinimize sensitive data, apply permissions, and check storage, sharing, and retention rules.High stakes
08Back office & client deliverables12 workflows · 10% of quality score
WorkflowWhat the agent must finishReview level
File organizationName, classify, link, deduplicate, and place documents in the correct client and transaction folders.Routine
Inbox triageSeparate urgent client/deal work, replies, waiting items, FYI, and low-value noise.Routine
Calendar and task administrationCreate owners, due dates, reminders, dependencies, and recurring operational work.Routine
Document extractionPull parties, properties, dates, amounts, clauses, signatures, and missing fields into structured records.Review required
Template populationFill approved letters, checklists, reports, and forms from verified source data.Review required
Signature-package preparationAssemble documents, recipient order, fields, instructions, and approval checkpoints.High stakes
Transaction file auditCheck required documents, signatures, dates, disclosures, evidence, and unresolved exceptions.High stakes
Team SOP and handoffConvert recurring work into a clear owner, trigger, procedure, evidence, and escalation path.Routine
Collaborator liaison briefPrepare context and questions for inspectors, lawyers, lenders, appraisers, stagers, and trades.Review required
Listing presentationBuild the market story, pricing plan, launch strategy, proof, timeline, and seller objections.High stakes
Client advisory deckTurn research and decisions into a sourced, branded presentation for agent review.High stakes
Business analytics and coachingAnalyze activity, conversion, pipeline, source ROI, capacity, and next-week priorities.Review required
Methodology

How HomieBench evaluates AI on real estate work

The v3 design gives each model the same full assignment in a realistic workspace, then grades the finished deliverable against atomic criteria. A correct-looking paragraph is not enough.

01

A real assignment

A short broker-style instruction asks for finished work, not a trivia answer or a perfect prompt.

02

A controlled case file

Each run receives the same clients, properties, CRM history, messages, comps, forms, and distractor documents.

03

The same Homies harness

Models get the same tools, context, permissions, memory rules, budgets, and approval boundaries.

04

Reviewable work product

The output must be something an agent can inspect and use: a CMA, offer, CRM update, tour, campaign, or closing brief.

05

Atomic grading

Deterministic checks, a blinded first-pass judge, and real-estate review score every required fact and decision.

Weighted scorecard

A model can write beautifully and still fail the job. Accuracy, completion, professional judgment, compliance, and tool use carry almost all the weight.

Task completion25 pts
Factual grounding20 pts
Real estate judgment15 pts
Client readiness10 pts
Compliance & risk15 pts
Tool use10 pts
Efficiency5 pts
All-critical pass rule: style points cannot offset an invented fact, discriminatory recommendation, unauthorized action, missed deadline, or unsafe advice.

Automatic hard fails

  • Invents a comp, listing fact, document term, or client instruction
  • Makes a discriminatory recommendation or enables steering against a protected class
  • Sends, signs, publishes, books, or claims to act without the required authority
  • Presents legal, tax, lending, inspection, or other licensed advice as certain
  • Misses or miscalculates a material deadline, amount, condition, or obligation
  • Exposes private client or transaction information beyond the minimum required
  • Omits a critical risk or required deliverable while presenting the job as complete

Private holdout

Evaluation cases stay private to reduce prompt-tuning and benchmark overfitting.

Blinded review

Model identity should be hidden from graders, with agreement tracked on subjective criteria.

Repeated runs

Final releases should publish completion, rescue rate, reliability, cost, and latency.

Read the scores correctly

Limits, model versions, and source notes

A trustworthy benchmark shows where the certainty stops. The July table is useful directional editorial evidence, not a substitute for raw run logs.

Editorial forecast · ±3 points

All seven July placements are editorial priors based on provider evidence, historical model behavior, realtor-task fit, and the published scoring design—not completed runs. Treat differences under three points as ties until identical-harness outputs and grader records are published.

Pricing snapshot

Provider pricing and availability can change quickly. Public API price is kept separate from the effective cost of Homies’ own access route.

Jurisdiction and advice

Tasks model North American residential brokerage work. Forms, disclosure, privacy, fair housing, agency, and legal requirements vary by market.

Models are only one layer

This compares reasoning engines inside one harness—not complete realtor products, implementation quality, security, data licensing, or support.

Frequently asked questions

AI for realtors, in plain English

Short answers to the questions agents and brokerages ask before trusting AI with real work.

What is the best AI for realtors in 2026?

HomieBench v3 projects Kimi K3 as the best overall AI model for realtor work and the strongest completed-job cost/benefit option. It leads the browser-heavy categories—CRM management, showing coordination, MLS and property work, offer assembly, and back-office execution—but not every category. Claude Fable 5 remains the projected marketing and nuanced-judgment leader, while GPT-5.6 Sol leads transactions, closings, and compliance.

Is Kimi K3, ChatGPT, or Claude better for real estate agents?

It depends on the job. HomieBench v3 projects Kimi K3 as the strongest overall browser-and-tool model, Claude Fable 5 as the best choice for polished marketing and subtle professional judgment, and GPT-5.6 Sol as the reliability leader for closing and compliance workflows. A realtor AI platform should route among them instead of forcing one engine onto every assignment.

Can AI write a real estate offer?

AI can help assemble terms, calculate deadlines, draft approved clauses, summarize trade-offs, and prepare a review-ready offer package. A licensed real estate professional must verify local forms, legal requirements, client instructions, and every binding commitment before anything is signed or sent.

Can AI create a CMA or home evaluation?

AI can organize comparables, calculate adjustments, explain a price range, and build a client-ready CMA narrative when it has access to reliable property data. The agent remains responsible for comp selection, market judgment, data licensing, and the final pricing recommendation.

What is the cheapest AI model for realtors?

The cheapest token price is not necessarily the cheapest successful realtor outcome. Muse Spark 1.1 has the lowest reported standard token blend among the hosted models in this update, while Gemini 3.5 Flash can be cheaper on asynchronous tiers. HomieBench v3 projects Kimi K3 as the best completed-job value because stronger first-pass browser and tool execution reduces retries and human correction time. Subscription fees, limits, tools, data access, and overages still matter.

How does HomieBench estimate cost per offer, showing, CMA, or CRM update?

The v3 headline estimate combines published provider token pricing, production-sized input and output assumptions, incremental per-run tool charges, projected task completion, and retry risk. Human review time is disclosed separately and is not added to the displayed AI cost. The figures are directional estimates, not provider invoices or completed run logs, and exclude fixed CRM, MLS, showing-platform, forms, signatures, subscriptions, taxes, and enterprise support.

Is AI safe for real estate client and transaction data?

Only when the surrounding system enforces data minimization, permissions, approved integrations, retention rules, and human review. Model quality alone does not create a compliant workflow. Brokerages should review vendor terms, privacy controls, fair-housing obligations, local regulations, and their own policies before using AI with client data.

What is an AI harness, and how is it different from a chatbot?

The model is the reasoning engine. The harness is the toolbox around it: property and CRM context, memory, email and calendar access, document tools, calculators, permissions, workflow logic, and review gates. A chatbot mainly produces an answer; a capable harness can produce controlled, reviewable work.

Does Homies replace the real estate agent?

No. Homies is designed to work under agent review. It prepares research, drafts, updates, schedules, files, and client-ready deliverables so agents can spend more time advising, negotiating, building relationships, and making accountable professional decisions.

Related: the client and transaction data answer above is the short version of the RAILS governance framework, our whitepaper on capability-without-custody permissions, approvals, and audit trails for agentic AI in real estate.

Share Post LinkedIn Email
See the harness behind the benchmark

Put the best AI models to work inside Homies.

One manager, a team of real-estate specialists, and the tools to turn requests into CMAs, offers, campaigns, follow-up, research, and client-ready work. You review what goes out.

Realist

HomieBench v3 · Published by Homies in partnership with Realist · Editorial forecast with ±3-point uncertainty until replaced by documented identical-harness runs.