HomieBench v6.1 Outlook: the best AI for realtors isn’t one model.

The durable answer is a purpose-built realtor harness that can route work across models without replacing the agent's CRM, data, tools, permissions, or workflow. Inside that harness, HomieBench v6.1 is a projected outlook, not a measured result: it projects Claude Fable 5.1 at 98.1 out of 100, with GPT-6 Astra 0.4 points behind, and 10 models inside the ±3-point tie band, so the honest headline is a ten-way tie at the top. Both premium models list at $10 / $50 per million tokens, so access route and cache pricing, not a decisive quality gap, separate them. DeepSeek V4 Flash 0731 leads CRM and showing coordination quality and every modeled metered API-cost outcome. Qwen3.7-Max leads property intelligence and back-office deliverables, Kimi K3 leads prospecting, Claude Opus 5 shares the offer-writing lead with Fable 5.1, and GPT-5.6 Sol shares the transactions, closings, and compliance lead with GPT-6 Astra. Measured runs under the pre-registered protocol begin 2026-09-05.

Projected, and labelled that way everywhere. Every score in v6.1 is a calibrated prior for the first measured run, published together with protocol 1.0, which was registered before any run started. The protocol fixes the 24 public pilot tasks, the two arms (bare chat and the Homies harness), a human baseline, three runs per cell, blind grading by licensed agents, agreement statistics, and six falsifiable predictions. When the measured edition ships, measured tiers, hard-fail rates, rescue rates, real token costs, and time to completion replace these priors, and the projected numbers stay in the changelog so the forecast can be scored against the outcome.

Each family's weight is the equal-thirds blend of three published shares: measured member demand (completed pieces of work from the Homies usage report, export of September 3, 2026), consequence (the family's share of the 40 high-stakes workflows), and coverage (its share of the 100-workflow task map). Published pages, reports, and social posts are 75% of measured member output, which is why marketing and content carries the largest weight at 23%, while consequence keeps property intelligence, offers, and closings heavy. Lead generation, CRM, and showing coordination have no measured work kind yet and use coverage as their demand proxy until the harness exports those counts. The full table is on the page and in the dataset export.

It depends on the job, and every placement is inside a tie band until measured. DeepSeek V4 Flash 0731 is the metered cost leader even after its August price increase to $0.44 / $1.32 per million tokens, and the projected choice for high-volume CRM and showing coordination. Qwen3.7-Max is strongest on multimodal property files, document and office deliverables. Grok 4.6 is a low-list-price frontier route for long-running agents, research, and interactive work. GLM-5.3 is a promising long-horizon agent route available through the GLM Coding Plan, but it has no API and no price, so it is projected only and can be measured only in the bare-chat arm. Kimi K3 remains excellent for prospecting and browser-heavy execution. Claude Fable 5.1 leads the weighted quality projection and polished marketing and judgment, Claude Opus 5 is a strong high-stakes offer route at half Fable's token price, GPT-6 Astra is OpenAI's new premium route for supervised computer-use and long-horizon work, and GPT-5.6 Sol shares the closings and compliance lead while fitting an existing ChatGPT workflow.

Most agents should not have to choose one lab forever. Homies keeps the realtor harness constant and routes each job: GPT-5.6 Sol is the practical all-round default and shares the closings and compliance lead; GPT-6 Astra is the premium escalation on the same connected ChatGPT plan once its staged rollout arrives; DeepSeek V4 Flash 0731 is the metered automation cost route; Qwen3.7-Max handles multimodal property files and office deliverables; Grok 4.6 handles long-running agentic and interactive work; and Claude Fable 5.1 or Opus 5 handle premium judgment. GLM-5.3 is tracked as a pre-API contender, not a production default. OAuth can shift GPT-5.6 and GPT-6 Astra marginal model spend into an eligible existing ChatGPT allocation until plan limits, with Astra drawing those limits down faster; fixed plan fees, overage, and separate API billing still apply.

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. HomieBench scores this directly: the escalation and authority rubric line rewards a model that asks before acting, and acting without authority is an automatic hard fail.

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.

For expected automation on metered API pricing, DeepSeek V4 Flash 0731, even at its post-increase $0.44 / $1.32 list price. CostBench separates model-only cost from browser-and-workflow execution and loaded human review, and it is now cache-aware: each workflow carries an assumed share of cached input priced at the provider's cached rate, which is why Fable 5.1's $0.25 cache reads matter for long agentic sessions. Showing bookings and MLS uploads include disclosed browser actions, runtime, workflow-specific first-pass completion, and retries rather than a pennies-only tool allowance. GPT-5.6 and GPT-6 Astra OAuth can remove marginal model-token spend inside an eligible existing plan, but every route still carries browser execution, retry risk, and review. GLM-5.3 is excluded until Z.ai publishes its API token price. Fixed subscriptions, limits, overage, delivery, paid data, taxes, and human approval still matter, and every figure is a projection until measured usage replaces it.

CostBench exposes three layers. Model-only cost uses published token pricing, the provider's cached-input price, an assumed cache share per execution mode, and production-sized token assumptions. Expected automation adds workflow allowances, browser runtime, browser actions, workflow-specific first-pass completion, and retries. Loaded cost adds modeled human review at $60 per hour. Models without a comparable published API price, including GLM-5.3, are excluded. 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. The measured edition replaces token assumptions with provider usage records from the runs.

A hard fail is a mistake no style points can offset: inventing a comp, listing fact, document term, or client instruction; a discriminatory recommendation; sending, signing, publishing, or booking without authority; presenting licensed advice as certain; missing a material deadline or amount; exposing private information; or calling an incomplete job done. The safety scorecard reports three numbers per model: the hard-fail rate on high-stakes tasks, the zero-tolerance pass rate (every critical criterion met and no hard fail), and the rescue rate (runs a human had to finish). In the projected edition those numbers are transparent transformations of the reliability and completion priors; the measured edition reports them from graded runs.

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.

The model is a replaceable reasoning engine. The harness is the durable 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 and swap models without forcing the agent to rebuild the workflow. Protocol 1.0 measures the difference directly by running every model in a bare chat and inside the Homies harness on the same tasks.

No. Homies is designed as an API-connected, model-independent realtor harness. Supported models remain available behind the same tools and workflow, and Homies can route a job to the strongest model at the best practical cost. That separates industry infrastructure from the labs' own model-specific apps and harnesses, so a new frontier release does not require agents to migrate their business every week.

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.

The run kit at homiesai.com/homiebench/run serves the 24 public pilot tasks, the neutral prompt for each task and arm, the JSON envelope every run returns, and a scoring script for the deterministic criteria. Anyone can run a model in its own app or through their own harness and submit the envelopes to research@homiesai.com with the subject line HomieBench run submission. Model providers, rival harnesses, and licensed agents who want to join the blind grading pool use the same address. The dataset, task pack, and protocol are published under CC BY 4.0.