I ran Fable 5 for half day and the guardrails are the real story
Anthropic dropped Fable 5 and I immediately swapped it into our dev stack. We route everything through a single endpoint on zenmux, so the actual switch was changing one model string and watching the latency graphs. The good parts first because there are a lot of them. I threw a refactoring task at it: split a messy python service into modules, preserve the public api, and write tests that prove nothing broke. Fable 5 planned the whole thing, caught a circular dependency I did not mention, and verified the tests pass. With Opus 4.8 I usually have to nudge it a couple of times when it forgets to update the init file. Fable 5 just did it. Then I dumped our full codebase and asked it to find a race condition we had been hunting for a week. It traced the async flow, named the exact function, and described the interleaving that triggers the bug. That level of context digestion feels new. Opus is good at long context, but Fable 5 felt like it was actually reasoning across the whole window instead of pattern matching near the top. I also sent it a blurry dashboard screenshot from a client call and it rebuilt the html and echarts config including the tooltip formatting. My designer’s first words were "when did you learn front end." I did not. But here is the part nobody in the launch threads is talking about enough. It is slow. On high effort I am seeing 45 to 90 seconds for a single complex turn. Our latency graphs go from a flat green line to a jagged mess the moment Fable 5 traffic hits. And it is expensive. The same prompt that costs X on Opus 4.8 costs roughly 1.4 to 1.7X on Fable 5 because it generates more tokens and runs at a higher effort tier by default. It writes its own reasoning traces out loud and bills you for them. For research tasks the quality is worth it. For "rewrite this email" it is comically overpowered. The bigger issue is the silent fallback. Fable 5 is basically Mythos with guardrails. When your prompt touches cybersecurity, biology, chemistry, or