Greg IsenbergJune 13, 202624m

Claude Fable 5 is BANNED. What to do?

TL;DR

Cloud AI is rented access, not ownership: Greg frames the Fable 5 shutdown as a wake-up call that your workflow can vanish overnight because a provider, policy change, pricing shift, or government action controls the model.
Local models are now good enough for most work: He says the quality gap closed fast over the last 6 months, and that a model on a decent Mac or gaming GPU can handle roughly 80% of what many people use ChatGPT-style cloud tools for.
Start with the runtime, not the model: Greg recommends downloading Ollama or LM Studio first, with LM Studio as the easier on-ramp for non-technical users because it has a full interface and model browser.
Hardware fit matters more than hype: His rough map is 4B models for almost anything, 12B as the sweet spot for 16 GB RAM, 27B to 35B for 30 GB-plus Macs or dedicated GPUs, and 70B-plus for serious setups like an Nvidia DGX Spark with 128 GB unified memory.
Qwen 3, DeepSeek, Gemma, and Llama are the four families to know: He calls Qwen 3 and the 3.6 series the best all-around choice, flags DeepSeek for strong reasoning but 10 to 30 second pauses, praises Gemma for fitting in 16 GB RAM, and notes Llama's huge ecosystem of fine-tunes and tutorials.
The real business angle is privacy and resilience: His startup ideas focus on on-device AI for healthcare, legal, and finance, local-first versions of existing AI tools, air-gapped agents for sensitive operations, offline AI for no-internet environments, and "resilience as a service" as insurance against cloud outages or bans.

The Breakdown

A Friday evening government letter allegedly wiped out "the most powerful AI model on the planet," and Greg Isenberg uses that shock to argue for a practical hedge: run local models you actually own. His case is simple: cloud is still best, but if 60 to 80 percent of your work can run on your desk, you need that generator-in-the-garage layer before your provider disappears, gets banned, or prices you out.