Your Next AI Subscription Shouldn’t Be ChatGPT 5.6 Or Fable 5. It Should Be Both.
Nate Jones’ comprehensive breakdown of model selection heuristics, the transition from raw benchmarks to model lineages, and the critical need for knowledge work harnesses.
Key Takeaways
- The Benchmark Fallacy: Raw benchmark scores (even private suites like Nate’s “Dingo” package or the “Agent/exam”) fail to capture the real-world operational utility of frontier models.
- Model Families & Resemblance: Models should be treated as families with distinct temperaments:
- OpenAI 5.x Family (Soul, Terra): Task-specific reinforcement learning (RL) optimization, strong on explicit, long-running agentic coding, but lacks “big model smell” and reading between the lines.
- Anthropic Mythos/Fable Family (Fable 5): Deep pre-training, philosophical reasoning, exceptional front-end taste, and excels at high-level ambiguity.
- The Selection Heuristic: Do not choose models based on benchmarks. Look at your best work, analyze your thinking process, and choose the model family that accelerates that specific loop. When in doubt, go with the model that helps you with your hardest work.
- Knowledge Work Harness Gap: While engineers have highly ergonomic tools (Codex, Claude Code), non-technical knowledge workers are underserved by tools built with engineering biases (e.g., ChatGPT Work, Anthropic co-work). True knowledge work requires process-oriented, conversational harnesses.
- J Space: Anthropic’s study on how models computationally manipulate higher-order concepts while performing lower-order token prediction.
Related
- nate-b-jones (Creator)
- model-families (Concept)
- knowledge-work-harnesses (Concept)
- fable-5 (Entity)
- chatgpt-5-6 (Entity)
- j-space (Entity)