Does your work cost less on Opus 5.5 now that its tokens are 20% cheaper? My 3-prompt guide shows which jobs do before you move them.

With Opus 5.5, I can give the model direction, make a correction, and keep developing the thing I wanted. That sounds ordinary until you’ve spent a session explaining the same requirement repeatedly and wondering why a system this capable is so difficult to work with.

The writing is much easier to steer. The visual work is extraordinary. I’m seeing efficiency improvements that make me think differently about how much I can get done with the intelligence I’m paying for.

There is a lot of intelligence available in these models. The amount we can use depends on how well we can work together. We have to communicate what matters, see enough of the work to judge it, and make changes without losing the decisions we’ve already made. When that gets easier, the next assignment we’re willing to try can get more ambitious.

That’s why I think model ergonomics deserves much more attention. By ergonomics, I mean the fit between the system and the person using it: the effort required to direct it, understand its work, correct it, and decide what to do next. It’s a useful way to understand what I’m enjoying about Opus 5.5. It also connects this release to the much larger story of AI helping build AI, faster development, and companies trying to understand what customers actually want from frontier intelligence.

Sonnet 5.5 just landed, and Anthropic cooked again. I’m still testing, but my early views reinforce a lot of what I’m arguing in the essay below.

Here’s what’s inside:

• Sonnet 5.5 points the same direction. What Anthropic’s newest release says about efficiency per task, and why it may shape how we experience compute availability.

• Why steering decides what you’ll delegate. How corrections that hold, even across connected problems, turn small safe requests into bigger assignments.

• The labs are working through this relationship themselves. What Anthropic and OpenAI report about AI-assisted development, and why faster releases have to preserve what users already learned.

• The business depends on people wanting to keep delegating. Why repeat use is the commercial story to watch as Anthropic reportedly heads toward an IPO.

• The real assignment test guide. Three prompts that turn your own recurring work into a test you can rerun every time a new model ships.

Start with the release that landed while I was writing this.