NewsOpinion

Fable 5 Is Back But With Some BIG Restrictions

Paul J Lipsky

Fable 5 AI model has been re-released by Anthropic but with significant restrictions: limited to 50% of weekly usage until July 7th, after which it requires paid credits, and enhanced safety classifiers that block more routine coding and cybersecurity tasks. The speaker discusses real-world usage results and the broader implications for AI industry gatekeeping.

Summary

Fable 5, Anthropic's most advanced AI model, has returned to availability but with multiple substantial limitations that disappoint users. The model is now restricted to only 50% of weekly usage limits through July 7th, at which point users must pay additional usage credits to continue accessing it. This represents a significant reduction from the initial Fable 5 release, when users had access to their full limits for several weeks before transitioning to paid credits. To release Fable 5 publicly, Anthropic implemented increased safeguards using AI classifiers that automatically detect and block requests for potentially harmful activities, particularly in cybersecurity and coding contexts. These classifiers have expanded safety margins beyond just blocking clearly harmful tasks, now capturing ambiguous requests and routine work. The speaker notes this is especially problematic for coders and cybersecurity professionals. However, during 30 minutes of testing, the speaker found that knowledge work tasks proceeded without restriction, and even some coding improvement requests worked fine. The only successful block trigger was asking about Chrome extension vulnerabilities. The speaker suggests strategically using Fable 5 only for the most complex tasks while relying on Opus and Sonnet for routine work to stay within usage limits. The speaker also frames this as part of a larger industry trend of government gatekeeping of state-of-the-art AI models, establishing precedent that limits public access to frontier AI capabilities.

Key Insights

  • Fable 5 is restricted to only 50% of weekly usage limits through July 7th, after which it transitions to paid usage credits, providing significantly less free access than when the model first launched.
  • Anthropic increased safety classifier margins to block not just clearly harmful requests but also ambiguous and routine tasks, particularly affecting coding and cybersecurity work.
  • The precedent of government scrutiny and gatekeeping of state-of-the-art models means future frontier AI models will face similar restrictions and limited public access.
  • In 30 minutes of testing, the speaker found that knowledge work tasks were not blocked, and even some coding improvement tasks worked, with only vulnerability assessment questions triggering the safety classifier.
  • Fable 5 is described as an incredible but overkill model for most tasks, suggesting a strategic approach of using it only for complex work while relying on Opus and Sonnet for routine tasks.

Topics

Fable 5 availability and restrictionsUsage limits and pricing changesAI safety classifiers and censorship concernsReal-world testing and performanceAI industry regulation and government gatekeeping

Transcript

[0:00] Well, Fable 5 is back, but there's a catch. Actually, there are several big catches that are going to upset a lot of people. I'm going to go over what those are, and we'll also look at what you can and cannot do with the new Fable 5 and all the restrictions that have been placed on it. I've been doing some quick testing around it, and have some helpful results for all of you. So, first, let's start with some good news. Fable 5 is back. If you open up Claude, you should see this now. Here on the desktop app, you may see this big message or on the model picker, you'll see Fable 5 in [0:32]…

Full transcript available for MurmurCast members

Sign Up to Access

More from Paul J Lipsky

Get AI summaries like this delivered to your inbox daily

Get AI summaries delivered to your inbox

MurmurCast summarizes your YouTube channels, podcasts, and newsletters into one daily email digest.