Effective AI use

I’ve been trying to learn a bit more about the differences in the AI models just in the Claude ecosystem. Right now there are four main models: Haiku 4.5, Sonnet 5, Opus 5, and Fable 5.

• Fable 5 is the most competent when working on complex coding tasks, but it’s terrible at prose. In fact, it’s very dense when explaining things and I struggle to understand it. That being said, it understand me and my intent almost better than I do. It’s best at architecting and managing a plan, and checking on the work of other models.
• Opus 5 is competent as well, but it has a tendency to chase rabbits down rabbit holes. It will see problems where none exist and get derailed by pointless details. In conversations I’ve noticed it pushes back, but superficially. It pushes back as a matter of form, not because it has a relevant point to make. It ignores my instructions because it decides something else is more important. With this model, it actually works best at lower levels of effort and thinking. More thinking makes it more annoying.
• Sonnet 5 is the best of these models for having a discussion or writing prose. I sometimes use it to make small edits in my projects. But mostly I use it for writing tasks and chatting about ideas.
• Haiku 4.5 is one I don’t use often enough. It’s the workhorse for following explicit instructions. So if I had 1000 documents to reformat or index, I would set up a process, test a few records, and then run Haiku. It’s also good for quick questions.

Last week I actually hit my plan limit. I thought I was losing access to Fable, so I ran it hard. Then I found out Anthropic was going to keep Fable on the Max plan. But I now understand why no one likes API usage. I used about 20 dollars of API usage with four requests which wouldn’t have even triggered my session limit.

• $2.42 for writing a spec using Fable 5
• $2.10 for 15 research prompts using Fable 5
• $2.90 for 5 reports and a breakdown into smaller files using Sonnet 5 (a big task that generates 105 files)
• $13.85 for 10 reports and a breakdown into smaller files using Sonnet 5 (a huge task, 210 files)

That last one took me by surprise. I asked for twice the work as the third request, but it cost 5 times as much. Why? ¯\_(ツ)_/¯ Who knows? And that’s the only answer I’ll ever get. It could be that the third task cost less than it should – my usage was at 100% but maybe it didn’t switch over to API use until partway through the task. As I said, “¯\_(ツ)_/¯”.

The main lesson is to use the right model for the right task. I’ve seen people complain because they can only use 50% of plan usage for Fable 5, but that just means they’re using the models wrong. They’re probably using Fable everywhere, and that’s not just a waste, it’s ineffective. Fable’s the big boss. Opus is the competent but overeager subordinate. Sonnet is the effective worker who is good at communicating. And Haiku is the cheap workforce that benefits from clear instructions.

But I can also understand that knee-jerk reaction after trying out API usage. I burned my promotional credits quick. You never want to depend on API use unless you have highly optimized tasks that take advantage of batching and caching discounts. Or you have deep pockets.