Had both Fable 5.1 and Astra write a project plan. Then had new instances of Fable and Astra compare plans and choose the better plan. Both chose Fable's plan. The project was to build a benchmark for a workerd feature I'm working on. Both agreed that Astra's plan was over-engineered. It spec'd out building a load generator from scratch as a node app. Fable checked what I had installed and proposed using vegeta or oha. Frankly, Astra's plan is also harder to read, over-specifying details with a lot of unnecessary "do not"s. They used similar token counts but Fable was about 2/3 the cost presumably due to charging much less for cache hits (major pricing change in 5.1 that people didn't seem to notice...). I also ran a head-to-head writing some C++ code a couple days ago and found Fable's code better-designed and better-documented. On code review, I had previously concluded Sol was better than Fable, and Astra is definitely better than Sol, able to discover even more, subtler problems. Of course, none of this is scientific, just my vibes based on a pretty small number of trials. I'll keep doing more tests but at present I am feeling like the status quo hasn't changed: Claude is better at writing (plans and code), GPT is better at reviewing/debugging. (And I'm getting really sick of having to re-evaluate this stuff every two months... TBH a "pause" might be nice for that reason alone.)