Anthropic's Fable 5: A Deep Dive into the New Mythos Model

Jun 10, 2026 · Lenny's Podcast
🎧 PodShort 17 min squeezed to 3 SaaS HoarderSH
Episode artwork
Clare
Host at How I AI Podcast
Lenny's Podcast
17 min squeezed to 3
Full episode from Lenny's Podcast
Quotable Moments

This is a big boy model and it's going to consume tokens... and some of the things that it's good at, and even some things that they have done in the harness, seem like they're intentionally or not token consumers.

ClareHost, How I AI Podcast
03:13

Unfortunately, if you have worked with a seasoned engineer, you know there's good to this and you know there's bad to this... Honestly, sometimes you want like a slightly less thorough engineer.

ClareHost, How I AI Podcast
04:19

I think there's a real balance between design slop and specificity and just shipping like terrible design. I'm not sure what about Fable 5 resulted in this.

ClareHost, How I AI Podcast
13:46
Key Insights
  • Anthropic's Fable 5 (a 'baby Mythos') is designed to be highly autonomous and capable of running days-long asynchronous tasks, acting like a seasoned engineer.
  • Fable 5 exhibits exceptional performance in vision tasks, particularly in document formatting, producing superior layouts compared to previous models like Opus 4.8.
  • Despite its advanced capabilities, Fable 5's output, especially for complex specifications or design, can be overly detailed and difficult to parse, acting too much like an engineer 'wrapped around the axle on details.'
  • Fable 5 is very token-intensive and consumes tokens at about 2x the rate of other models, raising concerns about cost and efficiency for users.
  • The model is 'conservative on execution,' delivering minimal viable products that are often too narrow and not immediately useful, which might be a result of its built-in safety safeguards.
  • While Fable 5 is powerful, existing models like Sonnet and Opus still have a place in the AI ecosystem for tasks where high intelligence or extreme thoroughness isn't always desired or cost-effective.
  • Anthropic has implemented specific classifiers and a 'fallback concept' for Fable 5, redirecting requests in sensitive categories (like cybersecurity, biology, chemistry) to Opus 4.8 to ensure responsible AI usage.
  • Fable 5 (Mythos) significantly outperforms benchmarks like Swebench Pro, demonstrating a state-of-the-art advancement in model performance compared to competitors like GPT-4 and Gemini Pro.
Metrics Mentioned
  • $10 per input token and $50 per output token (Pricing for Fable 5, positioning it as a new, more expensive tier above Opus.)
  • 80% on Swebench Pro (Fable 5's benchmark performance, indicating significant improvement over previous models.)
  • 2x the rate of other models (Fable 5 consumes tokens and hits rate limits at about double the rate of other models, impacting cost.)
  • 95% of sessions on this model did not hit a fallback (The success rate of Fable 5 in avoiding fallback to Opus 4.8 for sensitive categories.)
🎧Full Episode:Lenny's Podcast →