Fable 5.2 is something else...
Join this channel to get access to perks: https://www.youtube.com/channel/@aicodeking/join Councils are now available on PandaCode: https://github.com/aicodeking/pandacode In this video, I'll test Claude Fable 5.2, Anthropic's new top model, on all six KingBench 4 projects and grade every app with the same strict rubric as the rest of my chart. First, I go through the launch: the 20% price cut, fast mode, the new medium default effort in Claude Code, and the fine print behind Anthropic's benchmark table. Then I show how Fable actually works by going through its real sessions, and I go through the emulator, the Docker alternative, the room designer, the keyboard app, Pokémon and the Fitbit app one by one. After that, I look at the final chart, the real cost, how much of my $200 Ultra plan it used, and a free skill I made to fix its expensive habits. At the end, I show how PandaCode Councils and cross-model teams can make it even better, and I give my one-line verdict. -- Key Takeaways: 🚀 Claude Fable 5.2 scores 70% on KingBench 4 and takes first place, well ahead of Opus 5.5 at 57.8%. 💸 The API price drops 20% to $8 per million input tokens and $40 per million output, but that is still twice what Opus 5.5 costs. ⚠️ Claude Code now runs Fable at medium effort by default, so you need to set it to high to get results like these. 🧠 It researches the real docs before coding, and it even booted its own Linux VM to test the Docker alternative. 🏆 The room designer scored 9.2, the highest score any model has gotten on any KingBench 4 task. 🔐 The Docker alternative would have scored 7.7, but a container escape capped it at 3.0. 📊 Across the six apps it used about 212 million tokens, which comes to about $222 at API prices and about 3 times Opus on comparable tasks. 👥 The run used about 45% of my weekly Fable allowance on the $200 Ultra plan, which works out to roughly $500 of API usage per week. 🛠️ The free Fable-Cheap skill tells it to edit instead of rewriting files, wait for its own agents and test like an attacker. ✅ On the $200 plan it is absolutely worth it for your hardest projects, but on the API it is a specialist you call in, not your daily driver.
