You were lied to about Fable

The video starts by dispelling common negative claims circulating about Anthropic's Claude 3.5 Sonnet (Fable) model, such as it being terrible at coding, overly expensive, or 'nerfed.' The presenter, despite past disagreements with Anthropic, emphasizes that the model has impressed him and many friends, making the widespread negative takes frustrating and potentially preventing developers from recognizing its true power. He aims to clarify issues around cost, subscription availability (especially concerning its removal after July 7th), and performance, particularly coding capabilities.

The presenter addresses the misconception about Fable's performance, specifically the claim that it's being 'rerouted' from coding tasks. He explains that Anthropic's statement about falling back to Opus 4.8 for 'routine tasks like coding and debugging' was poorly worded. While fallbacks do occur for sensitive topics like cryptography or security-related tasks, they don't significantly impact general coding. He details Anthropic's two-stage safety classifier system: a cheap 'probe' that monitors internal model activations and, if sensitive sections are triggered, calls a more expensive classifier for deeper screening. This system aims to reduce jailbreaks without excessively increasing computational cost for every request. He stresses that the system is evolving and improving, advising users to try the model for their day-to-day work rather than rely on misleading benchmarks or misinterpretations of Anthropic's communication.

The video criticizes specific benchmarks used to claim Fable's 'nerfed' performance, calling them unreliable and noting that the benchmark creator has a history of posting inaccurate numbers. The presenter points out that the benchmark's tests involve terms that trigger the model's safety filters, leading to skewed results. He highlights inconsistencies in the benchmark's reasoning scores for other models, further undermining its credibility. He asserts that, in his experience, Fable remains an exceptionally smart and capable model.

The presenter discusses the changes in Fable's subscription availability, noting that Fable now has a dedicated weekly limit, capped at 50% of the total weekly usage. He explains that this restriction, and Fable's temporary removal from subscriptions after July 7th, is not an attempt to force users into higher tiers but a consequence of Anthropic's limited GPU capacity. The model's initial price was also significantly lower than planned due to increased GPU availability. This temporary period for Fable in subscriptions serves as a 'marketing experiment' and 'user research' for Anthropic to understand power user behavior, estimate future GPU needs, and gather data before enterprises heavily adopt the model.

The video shifts to practical advice on reducing costs and maximizing Fable's usage. The primary recommendation is to avoid 'X high' and 'Max' effort settings, as they massively increase usage without significant quality improvement. The 'High' setting is recommended for most tasks. The presenter shares his strategy of using Fable as a manager of 'sub-agents,' particularly by teaching Claude to route tasks to Codex. This is beneficial because Codex has generous usage limits and excels at tasks Claude struggles with, such as computer use involving screenshots and large token-heavy operations like processing PDFs or auditing codebases. This agent-orchestration approach allows him to achieve substantial work (e.g., merging 15+ pull requests) within his $200 monthly subscription, significantly reducing overall spend.

The video concludes by summarizing the key takeaways: reduce costs by using Fable to manage cheaper sub-agents (like Codex) for token-intensive tasks, never use 'X high' or 'Max' effort settings, understand that Fable's current subscription restrictions are a temporary experiment driven by capacity, and recognize that claims of 'nerfed' performance are largely baseless. The presenter announces an upcoming video detailing his specific workflows, skill writing, and system prompt changes for maximizing Fable's utility, including his 'chaotic vibe proxy' for rerouting across different accounts.