Yesterday, we covered a widely shared argument: the differentiation in AI has migrated from the model to the harness. Get the workflow, the context window, and the eval loop right, and the underlying model becomes a commodity you swap freely. The argument had real evidence behind it. It also has a flaw that a counter-thesis published on X on June 10 identifies precisely.
The counter-position, circulating from an analyst writing on X, is not a soft disagreement. It is a structural rebuttal. The argument: a harness built on top of a model you do not own is a moat built on rented land.
The mechanism is not hypothetical. Fable 5 is the live example. When Anthropic silently degraded Fable 5’s performance for certain API tiers, builders who had invested months into harness optimization woke up to degraded outputs they could not explain and could not fix. The supplier had repriced capability without changing the listed API price. That is the specific risk harness-only defensibility ignores: the model owner can restrict access, reprice the API, or reclaim capability behind a tier boundary at any time, on their schedule, with no obligation to your product roadmap.
The harness-camp response, anticipated in the piece, is the Cursor example. Cursor moved its coding agent from a ranking outside the top 30 to the top 5 by changing the harness, not the model. The counter-thesis accepts this. The harness matters. The argument is whether harness-only is enough for durable defensibility.
The answer, per the June 10 analysis, is no, and the reason is replicability. A competitor with the same API access can replicate a harness. What they cannot replicate is a model that has been co-designed with your proprietary data, your evaluation loop, and your specific workflow so that all four compound together. The model, the harness, the workflow, and the eval loop are not separate stack layers you swap independently. They are co-design surfaces. Changing one changes the optimal shape of the others. The strongest systems tune all four together, and that full-loop compound is what cannot be purchased through an API call.
This is the synthesis both camps are circling toward, even if neither states it directly. Everyone agrees the model alone is not the moat. A closed-weight frontier model behind an API, unconnected to your data or your eval loop, does not compound. The disagreement is whether the harness alone closes the gap. The June 10 analysis says it does not, because a harness-only position leaves the most critical variable outside your control.
The connection to lab-IPO valuation is direct and has not been made explicitly enough in the mainstream coverage. If moats need models, then the companies that own the model layer and co-design it with the application loop are worth structurally more than pure-API resellers. This is not an abstract claim. It is the investment thesis underneath the Anthropic, OpenAI, and xAI valuations. The market is pricing the full-loop owners at a premium precisely because the harness-only camp remains dependent on them.
The honest synthesis is narrower than either camp claims. Harness quality predicts near-term performance; model co-design predicts long-term defensibility. A team that has neither the resources nor the data to co-design a model should still invest heavily in the harness. But they should not confuse near-term performance with a structural moat, because the supplier can adjust the ground beneath them without warning.
If you are building a product on a single frontier API and your differentiation story is entirely about the harness, run a scenario where that API reprices at two times current rates or restricts your tier within the next twelve months. If the product survives that scenario, you have a real business. If it does not, you have a harness on rented land.
Commentary circulating on X from analyst writing at @sahar__zadeh, published June 10, 2026.
