三周实测 Astra 与 Fable 后，作者预告将测 GPT-6 Sol
长文对比 Astra 与 Fable 的编码表现，结尾表示等 GPT-6 Sol 上线后再补一份对比评测。
After 3 weeks of working with Astra and Fable every day, I think Fable is my favorite for almost everything complex, except infra and anything that requires computer use. Astra was honestly a nightmare to work with. It can create miracles and do things previous models simply couldn't, but I couldn't trust it with my code. It constantly changed things I never asked it to change. I had to add a ton of guardrails, and even then, it would revert things, modify unrelated parts of the codebase, or add things I never requested. Working with Fable, on the other hand, has been magical. Even token-wise, it didn't use nearly as much as Astra. The thing that really changed my mind was one bug I couldn't fix. Astra spent an entire night on it and burned through around 70% of my $200 Pro tokens without solving it. The next day, I reverted everything Astra had done, gave the original problem to Fable, and asked it to solve it from first principles. 30 minutes later, the problem was solved. That doesn't mean Astra is dumb. Far from it. I think the problem is that when things get extremely complex, Astra can do way more than you asked and start branching out in every direction, while Fable stays focused on the actual problem. I'd still use both. I can finally see what each one is strong at. And now with Opus 5.5, I've only spent about an hour with it so far, but it performed very similarly to Fable, which makes me really happy with it. I'll report back once I test GPT-6 Sol and see how it compares.