# The Model is Not the Moat

> AI costs halve every few months and the leading models have converged, so your edge lies in the context and harness you own, not the model.

**Published:** 5 October 2026

**Source:** [https://www.elsewhen.com/blog/the-model-is-not-the-moat/](https://www.elsewhen.com/blog/the-model-is-not-the-moat/)

---

Every great technology eventually becomes cheap, but artificial intelligence appears to be in an unseemly hurry about it. In early 2025, getting one of OpenAI's best models to answer a hard science question cost about 30 cents; eighteen months later, a newer model was answering just as well for less than a tenth of a cent, a fall so steep that, if applied to air travel, it would put a first-class seat to New York at little more than the Tube fare out to Heathrow. 

[Epoch AI](https://epoch.ai/publications/the-plunging-price-of-thought), an independent research institute that tracks the progress and economics of AI, found that the cost of AI at any given level of ability has roughly halved every three months since 2023, a rate of decline that electricity, the computer and every other world-changing technology before it never came close to matching.

So whatever edge a frontier model gives the business that buys it is therefore gone within a few quarters, which is why the model cannot be the moat. What sets one business apart from another is the context it brings to that intelligence, and the question that matters is who owns that context and the harness that turns it into action.

## Intelligence is now a commodity

AI has been shaped by the pace of its releases, and whenever one model explodes onto the scene another follows like thunder after lightning, a cycle that has closed the gaps between the labs rather than widened them. [Stanford's 2026 AI Index](https://c3.unu.edu/blog/2026-stanford-ai-index-report-takeaways) found that by March, models from Anthropic, xAI, Google, OpenAI, Alibaba and DeepSeek sat within 25 Elo points of one another on the Arena leaderboard, which is close enough that for most enterprise work the choice between them has become a procurement decision rather than a strategic one. Buyers are behaving accordingly: open-weight models ran 13% of the tokens routed through [Vercel's AI Gateway](https://vercel.com/blog/ai-gateway-production-index-september-2026) in April and 56% in August, the first month they carried the majority.

Researchers have since shown how much of an agent's performance lives outside the model. In August, [Prime Intellect](https://arxiv.org/html/2608.23552v1) reported that its open-source Prime Agent harness raised the best score on ARC-AGI-3, a benchmark built so that memorisation cannot fake a result, from 30% to 95.5% by its own measure, using Anthropic's Opus 5, a model any enterprise can already buy.

Enterprise leaders are saying the same thing from the other end. Redis's [State of Context Engineering](https://redis.io/resources/state-of-context-engineering-2026/) report, published last month, found that 73% of IT and AI infrastructure leaders think agents fail more often because of broken context than broken models, yet only 4% have built context systems that improve with use. The models, in other words, are not the problem, the wiring is.

## What a harness does

A harness is the software that turns a model into an agent that can do a job. The model reasons and writes, while the harness decides what it sees, which tools it can use, what it remembers, what it is allowed to do and how people check its work.

For an enterprise, a good harness is where a general-purpose model learns how one particular business works: what counts as a priority customer, which approvals a refund needs, how the team writes a claims letter. Because that knowledge lives outside the model, it survives every model change, so each new release improves the whole system rather than forcing a rebuild, and every task the agents complete adds to the context the next one draws on. That accumulated context, unique to your business and workflows, is the moat.

## Who should own it

Everyone has access to the same models, so what sets a business apart is what wires them together. At Elsewhen, we build that wiring for clients and call it the Productivity Platform: business context that puts the client's data and domain knowledge in front of every agent, headless agents that can be called from anywhere in the stack, an orchestrator that keeps multi-step work on track, a Skills Library of reusable expertise and generative UI so people can see and steer the work. It runs in the client's own cloud, belongs to the client outright and works with any model, and because we start with one costly workflow and build the shared core once the return is proven, each agent after the first ships faster and cheaper than the one before it.

The labs will keep releasing better models and vendors will keep offering to hold everything around them. The model is not the moat. The moat is the context and expertise an enterprise builds into its own harness, which gets better every time it is used and protects no one but the business that owns it.

[See what we can build in four weeks. Get in touch.](https://www.elsewhen.com/contact)