Introducing Claude Fable 5: a Mythos-class model that we’ve made safe for general use.
Its capabilities exceed those of any model we’ve ever made generally available.
11/ The capability curve (Alex Albert)
SWE-Bench Verified: 62% (Sonnet 3.7) → 87% (Opus 4.7).
Each generation, *simplify* — strip scaffolding built around old models.
'Boost performance by removing instead of adding.'
https://t.co/WsOnrqjmKd
10/ Getting more out of the Claude Platform (Brad Abrams)
Four core techniques: prompt caching (90% input discount; cached tokens don't count vs. rate limits), Tool Search Tool (lazy-load tools), programmatic tool calling, compaction.
https://t.co/Dxw7OOJs7v
@bcherny Love those tips! For iOS apps verification I've used Argent from @swmansion and today it released to the public https://t.co/OmeJGLCg4G
Really improves Claude consistency!
The most illegal feeling thing I’ve ever built.
I can now feel if my SwiftUI haptic logic is correct without ever looking at a physical iPhone. Testing the Success vs Error patterns just by resting my hand on my Mac or Magic trackpad.
Let me know what you think!