Anthropic is going through the wringer. In the last fortnight, Claude Code leaked, Mythos “leaked”, and users have been left holding the bag.

Claude Code

Claude Code is widely regarded as the best “harness” attached to any large language model (LLM) and is the basis for their recently released "Cowork" agents. That matters because we know LLMs aren’t sufficient by themselves. The real power comes from the orchestration (the harness). Once that code was exposed, competitors can now catch up.The most popular explanation for why the leak occurred is that Anthropic is letting AI write and push its code with too little human oversight.

Mythos

The Mythos leak is a similar but different issue. Within AI research circles, it is understood that the major labs train giant models at the edge of what existing hardware can support. These systems are useful for discovering what is possible. They are often far less useful as products because they are expensive, cumbersome and difficult to serve at scale. Mythos appears to fit that pattern. Again, human error or lack of oversight led to this leak. Anthropic’s preferred framing is that this is a new step change in LLMs “too dangerous to release” which is just a convenient line for publicity.

Quotas

Now to the issue of access. Anthropic introduced peak-hour multipliers at the end of March (without communication), making usage limits burn faster during busy periods. This came after promotional generosity that briefly expanded quotas. There are hilarious posts by developers on $200 USD/month plans using 100% of their daily quota simply saying “hi” to their Claude Code.

These restrictions also coincide with the decision for Anthropic to disable support for users using OpenClaw with monthly plans. These restrictions stem from a deeper underlying issue of Anthropic’s insufficient access to GPUs.

This recent sequence of leaks and restrictions suggests a company under pressure and perhaps engaged in an overdue review of speed versus discipline.

The lesson I would take from this is: No one should build their business around a single AI provider. They’re working on borrowed time, limited capacity, fast, and loose principles. Diversify your intelligence to minimise the impact when crunch time comes.