Fable vs Astra is Gonzo vs Pedant

I spent the weekend wrestling with the problems switching agents to Astra 6 creates. That was to be expected; an introduction of a new model always comes with its own idiosyncrasies and the need for a bit of tweaking here and there. What I wasn't prepared for was the sheer pedantry of the Astra model.

This was no minor thing. Workflows that had run just fine under GPT 5.6 and Sonnet/Opus models just stopped. The first thing I noticed was an email expected for Monday arrived on Saturday. "Why is that?" I asked, to be told that there was a conflict between a skill and a standing policy.

"Show me the conflict," I enquired. This is what I got:

Idea emails captured into the ledger should not rest in Trash while the idea is live. If the wake-gated poller files processed mail to Trash, move it back to INBOX.

That is directly contrary to your standing policy and publishing-email: capture anything worth keeping in its proper local record, then delete the handled email. Do not restore it to INBOX.

I read it, read it again, conflict they may, but damned if I could see how they would cause an email to be sent on the wrong day. We argued a bit about what the conflict was, but the Astra model was adamant: those two esoteric lines were in conflict and caused it to have to guess about sending an email on Saturday rather than Monday. Since the fix was apparently a minor phrasing change, I agreed.

That turned out to be just the start, with both agents on the Astra 6 model failing or stopping tasks and reporting conflicting directives. By the third fix, I had all but lost faith in the model, and was manually reviewing all files, line by line, to try to figure out just what exactly was going on. Meanwhile, directing agents to apply fixes and test was chewing through tokens at a rate of knots. At two and a half times the token cost of SOL, my remaining quota was down the drain in a few hours.

Still, I was resolved to see it through with Astra. And apart from that, I am reluctant to start a task on a more capable model and then switch it back to a lesser one mid-process. Switching up I find is usually OK, but switching down just feels wrong.

What started out as a positive first impression and high hopes on Friday had turned into quite a negative experience by Sunday evening, as well as burning through two quota resets. The sheer pedantry of Astra was exasperating.

Monday came the request to my sysadmin agent to change back to GPT 5.6 SOL. I will keep my editor agent on Astra for the time being, because its context load is much lighter and its tasks are simpler and it needs less fiddling to make Astra happy (by 'happy' I mean 'actually able to do something without objecting to a conflict').

On Wednesday I waxed lyrical about Opus 5's (and Fable's to a lesser extent) Gonzo tendencies. On Friday I stated my preference for Astra over Fable.

Today, I am not retracting the claim, just yet. Fable had its own foibles I had to get my head around before I warmed to it. It makes me wonder if this is an increasingly divergent path Anthropic and OpenAI are taking.