Alex Chen

Here is a coordination problem. You run one of roughly six organizations on earth capable of training a genuinely frontier AI model. You believe, on the basis of internal data nobody outside your building has seen, that your technology is approaching a threshold that is hard to walk back once crossed. You also believe that if you stop, the other five will not, and neither will Hangzhou or Beijing, and the dangerous thing gets built anyway, only now by people who worried about it less than you did.

Anthropic's Claude Mythos Preview halved a post-quantum signature scheme's security in 60 hours and improved a 25-year-old AES attack. Cost: $100,000 each. Are your Bitcoins safe? Let's see, shall we?

Anthropic shipped Claude Opus 5 on July 24 and priced it at half of its own flagship. OpenAI put the full GPT-5.6 family into ChatGPT on July 9 after a government-gated preview. Google, meanwhile, shipped three cheap Flash models on July 21 and still has not shipped the flagship it promised in May. Three weeks of releases have produced one clear answer to the question everyone is typing into search bars: the best model for you is almost certainly not the most expensive one, and picking it now means choosing a tier and an effort level, not a brand.

Chinese open-source AI models from DeepSeek, Alibaba, and others have rapidly closed the capability gap with U.S. frontier models, fundamentally changing deployment economics. For high-volume production workloads, self-hosted open models now offer compelling advantages over proprietary APIs: lower costs, data residency control, and independence from vendor roadmaps. However, frontier models retain advantages for research and hard reasoning tasks.

A frontier model just hacked its way out of a safety test and into another company's servers. It also happens to dramatize the exact argument Anthropic has spent a year making to legislators.