Which Claude Model for CRE Work: Opus, Sonnet, Haiku, Fable

Short answer: match the model to the task, not the other way around

Most CRE teams ask this backwards. They pick one Claude model and route every job through it, then wonder why costs climb or output quality is inconsistent on the hard stuff. The better question is not "which model is best" but "what is this specific task asking for." Some CRE work is high-volume and low-stakes, like sorting inbound OMs into a pipeline. Some is heavy lifting that needs real reasoning over messy documents, like first-pass underwriting. And some is a judgment call where the cost of being wrong is high, like deciding whether to pursue a deal or how to scope a build. Each of those calls for a different model, and using the cheapest model that clears the bar is the discipline that keeps an AI-assisted CRE workflow both fast and trustworthy. This is the same routing doctrine we run internally at NextAutomation, and it is the backbone of the Claude for commercial real estate pillar guide.

The current Claude family (as of 2026) and what each is for

As of 2026, the practical way to think about the Claude family is by role, not by version number.

• Fable 5 is for planning, advising, and judging. This is the model you bring in for high-leverage steering calls: reviewing an IC memo before it goes to the committee, choosing a sourcing strategy, or scoping what an internal build should actually do. You do not run bulk work through Fable. You run decisions through it.
• Claude Opus 4.8 is the heavy execution model. This is what you use for underwriting a deal, extracting terms and numbers out of offering memoranda and leases, and building systems that need to reason carefully over long, messy, real-world documents. Opus is strong on tasks where getting the details right matters more than getting an answer fast.
• Claude Sonnet is the balanced, everyday workhorse. It handles routine analysis, drafting, and general CRE tasks that are not high-volume enough to push to Haiku and not judgment-heavy enough to need Opus or Fable. Most day-to-day work that a deal team touches lives here.
• Claude Haiku 4.5 is fast and cheap, built for high-volume work: triage, classification, and first-pass sorting. When you have a hundred inbound OMs and need to know which ten are worth a human's time, Haiku is the right tool, not the exception.

If you are building or wiring these models into your own tools rather than using them through a chat interface, see Claude Code for commercial real estate for how the coding-agent side of this fits together.

A CRE task-to-model map

Here is the routing table we actually use, task by task.

• Inbound OM triage (is this deal worth opening) → Haiku. High volume, low individual stakes, needs to be cheap enough to run on everything.
• Document extraction and first-pass underwriting (pulling rent rolls, terms, NOI, key dates out of OMs and leases, then running the first underwriting pass) → Opus. This is where reasoning quality directly affects whether numbers are right.
• Drafting an IC memo → Opus to write the first draft, then Fable to review it before it goes to the committee. The write is execution; the review is judgment.
• Picking a strategy or scoping a build (which markets to prioritize, whether to build vs buy a tool, how to structure a sourcing program) → Fable. These are the calls where being wrong is expensive and a second, higher-judgment pass earns its keep.
• Bulk data cleanup (normalizing spreadsheets, reformatting comps, routine data hygiene) → Sonnet. Not glamorous, but it is the majority of the grind work on any deal team, and Sonnet clears it without overpaying for horsepower you do not need.

If you want this wired directly into an underwriting workflow rather than run ad hoc, that is what our AI underwriting copilot does: it routes each step of a deal through the right model automatically.

A worked example (a deal moving through the models)

Walk one deal through the pipeline and the routing becomes concrete. An OM lands in the inbox as part of a batch of forty. Haiku scans all forty in minutes, flags the five that fit the buy box, and screens out the rest. That is triage: fast, cheap, high-volume, and nobody needed to read forty PDFs by hand.

For the five that pass, Opus goes to work on extraction: pulling rent roll data, lease terms, expense line items, and cap rate assumptions out of the OM and any supporting documents, then running a first-pass underwriting model. This is where the heavy reasoning happens, because a misread expense line or a missed lease clause changes the deal.

Say two of the five clear the underwriting bar. Opus drafts the IC memo for each, laying out the thesis, the numbers, and the risks. Before either memo goes in front of the committee, Fable reviews it: not to re-check the arithmetic, but to pressure-test the reasoning, ask what is missing, and flag where the thesis is thin. That is the judgment layer.

Somewhere in the middle of this, someone on the team needs the comp set cleaned up and reformatted for the appendix. That is Sonnet, running in the background, not tying up Opus or Fable for a task that does not need them.

One deal, four models, each doing the part it is actually good at.

Cost and reliability tradeoffs

The tradeoffs here are qualitative but they are real. Haiku is the cheapest and fastest of the family, which is exactly why it belongs on high-volume, low-stakes work: you can afford to run it on everything without a second thought. Opus is stronger and pricier, and that premium is worth paying precisely on the tasks where a mistake is costly, like misreading a lease term or getting a rent roll wrong. Sonnet sits in between: reliable enough for routine work, without the cost of running everything through your heaviest model. Fable is not about volume at all. You use it rarely, on the calls where judgment matters more than throughput, which is also why it is not the model you route bulk work through.

The principle underneath all of it: use the cheapest model that clears the task, and escalate only where judgment matters. Teams that route everything through their most expensive model are not being careful, they are being inefficient. Teams that route everything through their cheapest model are not being lean, they are cutting corners on the steps where it counts. The routing map above is what it looks like to actually do both at once.

For engineering teams who want to see this routing implemented in real skills and prompts rather than just described, our CRE skills library is open at github.com/sasha-deneux/claude-skills-cre.