There is a wide selection of highly capable models out there, much cheaper than a ChatGPT or Copilot subscription. If you've read Niko's article this month, you've heard about Kimi K3, an excellent Chinese open-weight model that recently beat Claude Fable 5 on a leading coding benchmark. There are dozens of open source models specializing in various tasks, costing a fraction of what you would pay Anthropic or OpenAI. So why is it still worth paying for a Claude subscription?
The answer is simple: the harness.
You may have the best model in the world, but if it isn't managed properly, it will underperform spectacularly. An LLM's ability to complete tasks beyond writing text comes from what it is connected to, what instructions it follows, and what structures and guardrails it operates within. In other words, the harness. Like a strong racing horse, it won't win anything without the right tack.
What is a harness?
The model itself does one thing: it predicts the next word, over and over, until it has produced a response. Everything else, whether it can search your document store, call a calculator, retry when a step fails, remember what happened three steps ago, or check its own output before handing it to you, comes from the software wrapped around it. That wrapper is the harness. It decides what the model sees at each step, what tools it is allowed to use, how much of the conversation history stays in view, and what happens when something goes wrong.
This is why two people can use what is technically the same model and get completely different results. One is using it through a thin chat interface with no memory beyond the current conversation and no way to verify its own work. The other is using it through a purpose-built tool that breaks the task into steps, feeds it only the relevant documents at each step, and checks the output before returning it. Same brain, different equipment. The results are not remotely comparable.
Same model, very different results
This is not a hypothetical. Recent evaluations of coding agents have shown the same underlying model swinging by 30 or more percentage points on identical tasks, purely as a function of the harness wrapped around it. No change to the model. No fine-tuning. Just a better structured task, better managed context, and a verification step before the answer goes out the door.
Cost moves the same way. Two harnesses solving the same task with the same model can differ by three or four times in the number of tokens consumed for comparable quality, because one manages context tightly and the other dumps everything into the model's lap and hopes for the best. At meaningful scale, that difference is the gap between a manageable monthly bill and one that makes finance ask uncomfortable questions. Performance and cost are not a trade-off here. A well-built harness tends to improve both at once, because most of what it does is prevent the model from wasting effort on things it does not need to see.
You don't always need the expensive horse
This has a practical consequence worth sitting with. For a well-scoped, specialized task, such as extracting specific fields from a standard lease, flagging a covenant type, classifying documents by category, a smaller open-weight model wrapped in a well-designed harness will often match a frontier model in a thin one, at a fraction of the cost. The harness does the work that would otherwise require raw model capability to compensate for: narrowing the task, structuring the input, catching errors. You are not paying for intelligence you do not need.
None of this is an argument that frontier models like Fable are obsolete, quite the opposite. For genuinely novel problems, ambiguous questions without a clear structure, or reasoning that has to hold together across many unfamiliar steps, the extra capability of a frontier model earns its cost. The skill is knowing which category a given task falls into before defaulting to the most expensive option available, and increasingly, before defaulting to the same option for every task in your workflow.
What this looks like in real estate
You can see this distinction already in the tools built for real estate. A purpose-built lease abstraction tool and a general chat interface can run on the exact same underlying model and produce very different reliability on the same set of documents, because one has been built to scope the task, verify extracted fields, and flag anything it is unsure about, while the other has not. That's the difference a harness makes.
This also changes what is worth asking when you evaluate a new AI tool or vendor. "Which model do you use" tells you comparatively little. More useful questions are how the tool scopes the task, what it does when it is uncertain, and how it verifies its own output before showing it to you. It is also worth asking whether the task at hand genuinely needs a frontier model at all, or whether a narrower, cheaper model with the right harness would do the job just as well. That question, more than model choice, is where most of the controllable cost in an AI budget actually lives.
2026: the year of the harness?
If 2025 was the year the industry got excited about agents, models with access to tools, capable of taking actions rather than just producing text, 2026 looks like the year that excitement matured into something more structural. The conversation has moved from "can the model use tools" to "how do we build the environment it operates in so it performs reliably." That is the harness question, and it is increasingly a sustainability question too. A harness that scopes the task tightly, routes it to a model sized for the job, and avoids feeding the model more than it needs to see, uses less compute and less energy to arrive at the same answer, or a better one. Used this way, AI stops being a single, general-purpose tool you reach for by default and becomes something closer to a set of fitted tools, each matched to the job at hand. Efficient and customizable turn out to be the same idea. That is worth remembering the next time the instinct is to open the biggest, most expensive model for a task a smaller, well-harnessed one could handle just as well.
Created with
