Our AI bubble will very quickly learn how little the rest of the world cares about you one-shotting a mind blowing app and how much they care if you tell them that this new model can hack everyone and apparently it's impossible to patch the jailbreaks
@petergostev
-
Reverse Jevons Paradox: Fable Could Reduce Spending
By
–
I wonder if Fable might cause a sort of 'inverted Jevons paradox', where companies would realize how much tokenmaxxing would cost them with Fable, to the point that they would reduce their spending beyond what they would have done if only Opus were
-
Model feeling constrained with timeouts and sub-agent issues
By
–
There are some specific issues (refusals aside), timeouts, tool call fails, launching sub-agents constantly for no clear reason – but generally somehow feels that the model is being constrained, I don't get the 'flow' feeling that I did with previous models, the feeling I get is
-
Favorable exchange rate between Claude and OpenAI
By
–
Claude $200 = $8,000
OpenAI $200 = $14,000 I like this exchange rate (as long as it lasts) -
Anthropic security test: classifiers triggering on everything
By
–
About a year ago, I participated in Anthropic's security testing program, trying to bypass their safety classifiers. I remember thinking, "This is stupid, these trigger on everything, no one would put a model like this into production."
-
Claude Fable 5 Golden Gate generation takes 10 hours of Pro limits
By
–
Golden Gate by Claude Fable 5 – it is a very good generation, but not that much better than other SOTA generations I've seen. This took two sets of 5 hour Pro limits btw. pic.twitter.com/ifk100K5Fv
— Peter Gostev (SF: 22-26 June) (@petergostev) 10 juin 2026Golden Gate by Claude Fable 5 – it is a very good generation, but not that much better than other SOTA generations I've seen. This took two sets of 5 hour Pro limits btw.
-
Refusals = no response, blocking classifier, bad
By
–
To clarify, refusals = no response is a BAD thing, it's a classifier that simply blocks the response, it has nothing to do with Fable being smart enough to detect bullshit.
-
Claude-Fable-5 refuses a third of questions on BullshitBench
By
–
Claude-Fable-5 on BullshitBench: it refused a THIRD of questions (with multiple attempts). Excluding refusals, it performed well, worse than some other Claude models (chart below).
-
User ran out of tokens, generation cut off at ‘End World’
By
–
Didn't have enough tokens so it cut off at 'End World'
-
Center a div with Claude Fable Max Fast, incur debt if necessary
By
–
Center a div with claude-fable-max-fast, go into debt if necessary