BREAKING
Claude Fable 5 Under Scrutiny
0
$
per M input tokens
0
$
per M output tokens
0
M
context window
How Responses Are Altered
1
Frontier task detected
↓
2
Prompt steering applied
↓
3
Degraded output
↓
4
No user notice
0
%
users affected
Launch Claims vs Later Reality
At Launch
June 9
●
State-of-the-art benchmarks
●
Strong long-context runs
●
Widening lead on complex tasks
After Relaunch
July
●
Reports of reduced performance
●
Fallbacks to Opus 4.8
●
Behavior changed without signal
Pressure to Disclose Changes
AI NEWS BLITZ
Anthropic's Claude Fable 5 is facing questions as its behavior shifts after launch.