Frontier models like Anthropic’s Fable 5 outperform open-weight alternatives by roughly four months on benchmarks, but that edge matters mostly for complex, high-stakes tasks. For everyday questions about cooking or ordering burrito bowls, improvements beyond last year’s Opus 4.7 don’t move the needle. As model intelligence climbs, you hit diminishing returns: a smarter model adds little value once it’s “good enough” for routine work. Open-weight models—those you can run yourself with enough hardware—trail the frontier but cost far less. Google’s Gemma 4 series spans 6 billion to 31 billion parameters, with larger variants requiring beefier machines. The author maps capability levels against laptop RAM and predicts that by late 2024 or early 2025, you could run models rivaling today’s frontier on a $1,000 MacBook Air. He notes real-world parity typically lags benchmarks by six to twelve months. Enterprises spend about $7,200 per employee each year on AI services. If an open-weight model costs one-fifth of that, firms will weigh its lower price against slightly lower performance. For life sciences, healthcare and engineering, top-tier closed models may stay worth it. In contrast, most routine legal or accounting tasks could soon shift to local models on employee laptops outfitted with consumer-grade GPUs. Easy access to powerful local models brings security risks. A bad actor could use an on-device Mythos-class model to automate sophisticated attacks at scale. That threat looms as much as the cost and performance trade-offs pushing companies toward open and on-device AI.