Too powerful to ship

The frontier labs spent 2024 racing to build the most capable models they could. Now they're racing to figure out which capabilities they can't let anyone use.

Anthropic just made that tension explicit. TechCrunch reported that the company's new Claude Mythos Preview model, scoring 93.9% on SWE-bench Verified and 77.8% on SWE-bench Pro, will not be made generally available. Instead, Anthropic created Project Glasswing to deploy Mythos exclusively to 12 companies — Apple, Microsoft, CrowdStrike, and nine others — for defensive cybersecurity. The model has already found thousands of zero-day vulnerabilities across every major operating system and web browser. Anthropic is backing the programme with $100M in usage credits and $4M in donations to open-source security organisations. When a lab builds a model and immediately concludes that its paying customers can't have it, the capability overhang has become the product strategy.

That containment logic falls apart when you look east. OfficeChai reported that Z.AI's GLM-5.1, a 744-billion-parameter mixture-of-experts model, just topped SWE-bench Pro at 58.4 — beating Claude Opus 4.6 (57.3) and GPT-5.4 (57.7). It was trained entirely on roughly 100,000 Huawei Ascend 910B chips using the MindSpore framework. No Nvidia, no AMD, no American silicon whatsoever. And Z.AI released it under MIT licence with a 200K token context window. So Anthropic is restricting its best model to a vetted coalition while a Chinese lab publishes a competitive one for anyone to download and run. The asymmetry is stark.

The Western labs know it. The Japan Times reported that OpenAI, Anthropic, and Google are now sharing information through the Frontier Model Forum to detect adversarial distillation — the practice of repeatedly querying ChatGPT, Claude, or Gemini to train cheaper copycat models. Anthropic previously named DeepSeek, Moonshot, and MiniMax as having illegally copied Claude's capabilities this way. U.S. officials estimate the practice costs Silicon Valley labs billions annually. Three companies that compete on everything else decided this threat was worth collaborating on.

Even Meta is pulling back. Axios reported that its next-generation models — an LLM codenamed Avocado and a multimedia generator called Mango — will ship as proprietary products first, with open-source versions following later and missing key features for safety reasons. Under new AI chief Alexandr Wang, Meta is hedging: open enough for developer mindshare, closed where the biggest models carry real risk. A year ago Meta was the loudest voice for open-source AI. Now it's adopting the same tiered-access playbook as everyone else.

What this means if you're building

The practical problem here is real. If you're building products on top of frontier models, the capability you can access and the capability that exists are no longer the same thing. Anthropic's best model is locked behind a security coalition. Meta's best models will ship with features withheld. And the one frontier model you can actually download and run without restrictions was trained on Chinese hardware by a Chinese lab, which introduces its own set of dependencies.

I think we've crossed a line. The question for product teams used to be "when will models be good enough?" Now it's which capabilities your provider will actually let you use, and under what terms. The models have outrun the distribution model. Builders who plan around today's access levels without accounting for tomorrow's restrictions are building on sand.

---