Fable 5.2 and Jagged AGI
ONEUSEFULTHING.ORG
Earlier this week Anthropic delivered another major model release, Fable 5.2, just a couple of weeks after Opus 5.5. Anthropic independently versions each model track, so even though this is a lower version number than Opus, the Fable class is the flagship model. And yes, it smashed the benchmarks as they shared in their release blog post, and it now sits 5 points ahead of its nearest rival (Gemini 4 Carbon) on the Artificial Analysis Intelligence benchmark.
As I have mentioned a number of times recently, the benchmark score and core capabilities of these models are quite incredible, and often times, they offer far more power than we know what to do with (or can afford). What matters more is how they feel, how smart or intelligent they are. Specifically for software development, they can write incredible amounts of code, but can they engineer?
Now I haven’t tried Fable 5.2 yet, but I have been using Opus 5.5 extensively for a few weeks now and I am very impressed by its engineering capability. It makes sensible decisions, compromises and pushes back at times. It exhibits a level of judgement I’ve not seen in previous models.
This blog post, by Ethan Mollick, shares his experiences with Fable 5.2. The things he has built are wild, but of most interest are his observations around its “judgment, playfulness and cleverness”. I look forward to giving it a spin.
How I’m fighting the AI Brain Fry
SERGEYKUHNE.COM
As our adoption of agentic AI grows, we find ourselves in a situation where we are managing multiple long-running sessions in parallel. As a result, we are constantly context switching, answering questions and providing additional context as each agent session fights for our attention. It can be mentally exhausting, and simply quite confusing. I often find myself losing track of my various sessions!
Sergey shares his own technique for managing this. The most important thing is that everything lives in git, it is the source of truth rather than the agent session. From this, he runs agents in a queue, with items requiring human attention managed by a single interface. As a result, he doesn’t have to use his brain as a task manager.
I’ve not yet found the need to intentionally manage my parallel agent sessions, but as the mental burden increases, I might need to give this a try.
Goodbye Claude, Hello Pi
CALEB.SOFTWARE
Claude Code is incredibly popular, topping the ‘AI Developer Tool’ charts in every survey I have seen. But it is far from perfect. Anthropic are iterating and changing Claude Code at an incredible pace, which unfortunately means that this tool, which has become critical for many of us, is in a constant state of flux. It also has poor IDE integration and an increasingly bloated interface.
In this post Caleb outlines why he has switched to Pi, having been attracted by its simplicity, stability and flexibility. Rather than waiting for a vendor to build features around its own vision of how developers should work, Pi lets you shape the tool around your own workflows. Caleb describes building a custom extension in just 20 minutes to solve an integration problem he’d struggled with in Claude Code.
While Anthropic are packing any feature they can imagine into Claude Code, Pi shows us that sometimes, less really is more.