If It Learns Slowly, Stop Teaching It

In July, Claude changed its pricing: Fable 5 is no longer part of the Pro subscription — it runs on separately metered usage credits. I’m a standard Pro user, so my first instinct was to economize: daily tasks go to Opus 4.8, credits stay saved for the big jobs. Daily content production didn’t count as a big job.

The pipeline in question was built back when Fable was in open trial: interview-style writing. The AI asks questions round after round, pries my actual views on a topic out of me, then turns the material into a draft. The first piece it produced I was genuinely happy with — the AI’s writing and mine were mixed to the point where it didn’t read like template AI prose.

After the switch to Opus, the interviews were the first thing to go wrong. It loved showing off — explaining things through strange, elaborate metaphors. At first I thought it was harmless. By the time it produced an outline, I felt something I hadn’t felt before: the logic and the spine of the draft were simply a mess. Point at any one part and it seems defensible; step back and everything reads wrong.

My first reaction was to fix the rules. The pipeline has a rule file constraining how the model writes, so I added clauses: no showing off, no gratuitous metaphors. The metaphors didn’t decrease. It didn’t feel like disobedience; it felt like the model couldn’t help itself. It also kept reaching for opaque words, so I added another rule to the style file: plain language over jargon. Still wouldn’t hold. The worst single instance: asked to explain the difference between a specialized model and a general model given structured prompts, it stacked four metaphors in one breath — weights as “hands,” context as “sheet music,” the same model with a new prompt as “a violinist in a costume.” I later had other models audit the clauses I’d written. Verdict: the rules were mostly fine. Good rules, and a model that can’t execute them — the problem isn’t the rule file, it’s the model. If it learns slowly, stop teaching it.

The second project pulls market data to support a decision. It also started on Fable during the trial. Fable ships with a safety classifier that hands security-adjacent topics down to Opus, so this one got downgraded. Fine — my past experience with Opus doing labor was decent. For the research I also brought in GPT. My one takeaway: what a perfect pair. Opus scraped a huge pile of App Store and Apify data — Apify close to exhaustively — analyzed it, and concluded: nothing here is worth doing. GPT got far less data, a few dozen entries, burned a week of my credits writing a report, and came back with three viable directions; I asked it to dig deeper into those three (another week of credits), and its final conclusion matched Opus exactly: don’t bother.

Since what remained after collection was analysis, I handed all the gathered data to Fable. It didn’t write yet another report. It told me to ship the Apify scraper as a public tool and let the market answer. It didn’t make me money either. But in an environment where nobody knows what survives, taking a step beats standing still: get a result fast, validate, retro. That’s what actually helping with a decision looks like — as opposed to analyzing everything and telling me to lie down and do nothing.

My conclusion from these two projects: brain work — anything where the model has to hold the judgment — goes only to Fable, for now. And Opus, as long as the judgment stays with you, is still a very good worker.