Thorsten Ball compares AI coding agents, finding Fable cost $457 while deep^2 completed the task for $20

IMO sth that is a bit overlooked but will become far more important in the future. GPT is 10-20x more token+cost effective for ~similar outcome.

@thorstenball My sense with fable in cc is that it is very good at orchestrating sub agents, and could probably do a great job commanding many opus instances.

In curious how you’re thinking about sub agents in amp.

@__tosh GPT-5.5 in Amp: https://ampcode.com/news/gpt-5.5

@psufka It wins at PowerPoint.

@robrhyne Hmm, skeptical it's going to be that important. Because it's hard to know in advance when "the dumber model can do it"

@thorstenball what is deep^2?

I think the trick with Fable we'll find out is that for cost effectiveness, we'll rely on it as a brilliant orchestrator because it understands the assignment well and can break it down effectively to other agents. This is because it has a very in-depth understanding of the problem, or at least, in my experience, it has actually understood what I was trying to say in a scenario where I was having difficulty articulating and Codex was thrashing for days so far. And Fable just got it immediately. Not just that, but because it did so, it could translate it in a way that would dramatically push the chances of success for a cheaper model like Haiku.

@beshr There's no Amp cost here

@thorstenball Deep is GPT 5.5?

@thorstenball How much of the cost is Amp's and how much is Fable?

@thorstenball With all the different capabilities and even _personalities_ emerging, I can’t help but think that model routing is going become an incredibly important part of the stack.

@thorstenball which model are you guys using on deep^2?

@steipete Agree… Curious how much you’ve tested Fable. Anything it’s better at?

@steipete @grok why would this be true?, what gives gpt this edge, what are some capabilitites not shared publicly that you guess make gpt more effectif for similar outcome in point form, is this irrepricable

@steipete

@steipete the problem is the ant models are just more fun to work with

@thorstenball How "deep^2" mode differs from "deep"?

@thorstenball I think it’s even better (or worse depending on who is looking at it). A really small but capable model (e.g. qwen3.6-27B) can reliably solve any simple enough task. Any complex task can be broken down into small simple ones and orchestrated back to solution. Frontier is for lazy

@steipete except for designs and writing

@komu_wairagu GPT-5.5