Fable Returns, GPT 5.6 Hype, Coding has changed so much | Ep 19

Fable Returns From the Dead

The hosts revisit Fable after its return from a temporary shutdown. Adam compares the experience to his first time using Opus 4.5, noting it feels like the model just does more now, but only if you use it correctly. The consensus: Fable is not a co-pilot model for quick tasks, but something you deploy when you need serious thinking power.

Fable as the Orchestrator Model

Ray tested Fable on iOS development, specifically building an animated camera component. He watched the thinking traces show the model reasoning through iPhone constraints and polishing the final implementation. His verdict: use Fable as an architect to plan work, then hand off to faster models like Composer 2.5 to execute. Nathan found Fable excels at creative writing and in-depth analysis, though he only got a partial one-shot on his Slay the Spire mod problem.

The Cost Problem

The economics of Fable are brutal. Adam burned $120 in API tokens during a short session. Ray notes Fable costs $50 per million output tokens compared to Grok at $6. The hosts wonder how users will justify the cost for anything but high-value work, especially when cheaper models could potentially match Fable results with more scaffolding and engineering effort.

GPT 5.6 and the Cerebras Speed Boost

GPT 5.6 is rumored to release with Cerebras-class hardware, potentially delivering 10x faster tokens per second than 5.5. Adam is hyped about the speed but warns that faster tokens mean burning through subscription limits and API budgets at unprecedented rates. The model lineup will include Soul, Terra, and Luna variants, replacing the clearer nano and mini naming scheme.

When Speed Actually Matters

The hosts debate when LLM speed actually matters. Nathan frames it as a risk versus reward calculation. He uses Opus as a personal trainer because the risk of injury from a bad suggestion outweighs waiting longer, but uses Haiku for cooking questions where he can quickly verify the answer. Adam notes his mental model has shifted: for async tasks, he no longer cares if it takes 20 minutes as long as the quality is there.

Claude Goes Vertical

Claude is executing a verticalization strategy with Claude Code, Claude Design, and now Claude Science. Nathan argues that companies evaluating competition need to consider Anthropic as a potential threat to their vertical. The strategy also makes sense from a data collection angle: bespoke workflows in specific domains help Anthropic gather expertise to improve their models over time.

The Coming Battle Over Model Access

Ray raises concerns about model access and regulation, pointing to China's lack of medical data restrictions as a potential advantage in training specialized models. He mentions the Freedom of Intelligence movement, organized by Quinn from AMP and others, advocating for the right to run local models. The hosts acknowledge the tension between frontier model capabilities and government nervousness about what these systems can enable.