Opus 5.5 and model ergonomics: why steering AI well matters more than raw intelligence

By Nate (Nate's Substack) · September 30, 2026.

Nate presents his analysis of Claude Opus 5.5 around a concept he calls "model ergonomics": the fit between the AI system and the person using it — that is, the effort required to steer it, understand its work, correct it and decide the next step. According to the author, this dimension deserves far more attention than it usually gets when assessing a new model, because it determines how much of the available intelligence can actually be put to use in practice.

Nate himself offers a three-prompt guide so that readers can use their own recurring tasks to evaluate which jobs are worth running on this model.

In his personal experience with Opus 5.5, the author describes being able to give the model a direction, make a correction and keep developing the original idea without having to repeat the same requirement over and over — something he had found frustrating in earlier sessions despite the model's capability. He highlights that writing is far easier to steer and that the visual work strikes him as extraordinary, with efficiency gains that make him reconsider how much he can get done with the intelligence he pays for.

For Nate, when communicating what matters, reviewing the work and correcting it without losing the decisions already made becomes simpler, people dare to delegate more ambitious tasks. In other words, ergonomics is not a cosmetic detail: it directly determines the size and complexity of the assignments a user is willing to entrust to AI.

The author mentions that Sonnet 5.5 had just launched while he was writing this piece, and that his first tests with that model reinforce the central arguments of his essay on ergonomics.

The email previews the content of the full article (hosted on Substack's website), organized into several blocks: first, how the launch of Sonnet 5.5 points in the same direction as Opus 5.5 in terms of efficiency per task, and what that might imply about compute availability. Second, why the ability to "steer" the model — getting corrections to stick even on interconnected problems — is what turns small, safe requests into bigger assignments. Third, how the labs themselves (Anthropic and OpenAI) are managing this relationship with their users, and what both companies report about AI-assisted development, underscoring that shipping models faster makes it necessary to preserve what users have already learned to do with previous versions. Fourth, a business argument: commercial success depends on people wanting to keep delegating tasks, and repeat usage is the key metric to watch, at a time when Anthropic is reported to be moving toward a possible IPO. Finally, the article offers what Nate calls "the real-assignment test guide": three prompts designed to turn the reader's own recurring work into a test that can be repeated every time a new model comes out.

The email is mainly an introduction and table of contents inviting readers to read the full article on Nate's Substack; the specific details of each section (exact figures, quotes from Anthropic or OpenAI, or the literal text of the three prompts) are not included in the body of the message.