{
  "id": 2858949,
  "title": "Fable and the End of the Free Lunch",
  "url": "https://urgent.news/2026/08/23/fable-and-the-end-of-the-free-lunch",
  "topic": "ai",
  "section": "AI",
  "published": "2026-08-23T19:06:09.000Z",
  "source": {
    "name": "Hacker News",
    "slug": "hacker-news",
    "url": "https://www.dbreunig.com/2026/08/23/fable-the-end-of-moore-s-law.html"
  },
  "original_language": "en",
  "account": "In the weeks following the release of Fable, a thought struck me: the era of free, high-performance models was coming to an end. When Moore's Law was in full swing, it was unnecessary to obsessively optimize code, as a new CPU would arrive every 18 months to double performance. Herb Sutter coined this phenomenon \"the free lunch\" in a seminal essay. However, as Moore's Law slowed in the mid-2000s, with single-threaded performance stagnating, we were forced to reconsider parallelization, architecture, and memory locality.\n\nPrior to Fable, it seemed wasteful to invest too much effort into refining coding harnesses or context strategies, as a new model would soon arrive at the same price or even cheaper, addressing most of our issues. But Fable's release changed everything. It was an incredible model, yet its cost was astronomical, and Opus, as well as earlier models such as 5.6, K3, and GLM, proved more than sufficient for most coding tasks. Consequently, we began to ponder where to allocate our efforts.\n\nThe release of GLM 5.2, just a week after Fable, was a turning point. It offered a solution at roughly 1/9th the cost of Fable (and nearly 1/5th the cost of Opus 5). While GLM may not match Fable's quality for all tasks, it is often more than adequate for routine coding. Moreover, when provided with excellent context, GLM has proven more than sufficient for many applications. In fact, I frequently use Fable to meticulously design and refine a plan before delegating the task to GLM. While some argue that decreasing inference prices will eventually lead us back to utilizing the largest models, I remain skeptical. These same gains will benefit smaller models such as K3 and Qwen, and as we continue to develop better harnesses, it will become easier to provide weaker models with sufficient context for optimal performance. Furthermore, Fable's access controls, dynamic degradation, and required data retention have alarmed various companies and even nations, prompting them to reconsider where they send their traces and where they acquire their tokens.",
  "summary": null,
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}