For years, the contest in artificial intelligence was framed as a race to build the smartest model. The companies with the largest computing clusters, the deepest research teams and the most advanced proprietary systems appeared destined to control the market.
Then, in June, Anthropic abruptly disabled access to Claude Fable 5 and Claude Mythos 5.
The company said the United States government had ordered it to suspend access to the models for every foreign national, including foreign employees working inside the United States. Because Anthropic could not practically separate permitted users from restricted ones across its systems, it withdrew the models from all customers. Other Claude models remained available, but two of the company’s most capable systems had effectively disappeared days after their launch.
The government cited national-security concerns related to the models’ cybersecurity capabilities. Anthropic disputed the severity of the identified risk, saying the demonstrated technique uncovered only a small number of previously known, minor vulnerabilities and that other publicly available models could perform similar work.
Whatever the merits of that dispute, the episode revealed something larger than a fight between an AI company and Washington. It showed that a proprietary model is not merely software purchased by a customer. It is access rented from a company operating under the laws, commercial priorities and political pressures of its home country.
For governments, banks, defence organisations, research laboratories and global businesses, that dependency is becoming increasingly difficult to ignore.
A Powerful Model Is of Limited Value if Access Can Vanish
Cloud-based proprietary AI offers obvious advantages. It is easy to deploy, requires little infrastructure and gives customers immediate access to systems maintained by some of the world’s best-funded laboratories.
But convenience has often concealed a structural weakness: the customer does not control the model.
The provider controls availability, pricing, usage policies, safety filters, data-retention requirements and regional access. A company can withdraw a feature, discontinue a model, change contractual terms or block a country. A government can impose export controls. A geopolitical dispute can interrupt services that have already been integrated into critical operations.
The Fable episode turns that abstract risk into an operational one. Anthropic had described Fable 5 as a model capable of advanced software engineering, analytical work, visual reasoning and long-running autonomous tasks. Mythos 5, built on the same underlying system with fewer cybersecurity restrictions for selected users, was being made available through a controlled programme involving the United States government.
Within three days of their public unveiling, access was suspended.
For an individual experimenting with an AI chatbot, such disruption may be inconvenient. For an organisation that has built its code review, research, fraud analysis or intelligence workflows around a particular model, it can be far more serious.
This is where open models offer a different proposition. Once an organisation lawfully downloads the weights and deploys them on its own infrastructure, continued use no longer depends on an external application programming interface remaining available. The model can operate inside a private cloud, a national data centre or an isolated network. It can be customised for local languages, internal documents or specialised tasks. Its operation is controlled by the organisation rather than by a distant provider.
The argument is not that open models are immune from regulation. Governments can still regulate chips, deployment and distribution. But ownership of the model weights gives institutions a degree of continuity and autonomy that an API account cannot provide.
GLM-5.2 Shows How Narrow the Performance Divide Has Become
The case for open AI would be weaker if organisations had to accept dramatically inferior performance in exchange for control. Increasingly, they do not.
Z.ai’s newly released GLM-5.2 is an instructive example. The company has released the model under the permissive MIT licence, with a one-million-token context window and support for local deployment through common inference frameworks.
According to Z.ai’s own evaluations, GLM-5.2 is competitive with prominent proprietary systems across several reasoning, coding and agentic benchmarks. It does not defeat every closed model on every test. On some difficult software-engineering evaluations, Claude remains ahead. On others, GLM-5.2 approaches or exceeds results reported for models from OpenAI, Google and Anthropic.
Those figures should be read cautiously. Benchmark results are often produced by model developers, depend heavily on testing conditions and do not guarantee performance inside a particular organisation. Yet the broader direction is difficult to dismiss.
Stanford University’s 2026 AI Index found that the leading closed model still held an advantage over the strongest open-weight model as of March 2026, but the gap was only about 3.3 percent on the Arena leaderboard. That is a meaningful lead, but it is not the gulf that once separated open systems from frontier laboratories.
For many organisations, the relevant question is no longer whether an open model is the absolute best in the world. It is whether the model is sufficiently capable for a defined task while offering better privacy, control, customisation and continuity.
A bank may prefer a slightly weaker model that can run entirely inside its own environment. A government may value a system that cannot be disabled by a foreign vendor. A hospital may prioritise keeping sensitive records within a controlled network. A defence organisation may care more about auditability and sovereign deployment than a marginal benchmark advantage.
In such environments, “good enough and controllable” can be more valuable than “best but revocable.”
The Future Is Likely to Be Open, but Not Automatically Safe
The movement toward open models should not be romanticised.
Releasing powerful model weights can make useful capabilities available to researchers, startups and smaller countries. It can also give malicious actors systems that cannot be monitored or centrally restricted. Once an open model is downloaded, its safeguards can be modified, removed or replaced. The same flexibility that allows a hospital to adapt a model for clinical use can allow a criminal group to adapt it for fraud or cyber operations.
There is also a terminology problem. Many systems described as “open source” are more accurately called open-weight models. Their weights may be downloadable while their training data, development code or complete training process remain undisclosed. Licences may restrict commercial use or certain categories of deployment. Genuine openness exists on a spectrum.
GLM-5.2 is notable because Z.ai says it is being released under the MIT licence, one of the more permissive software licences. But even a permissively licensed model is not costless. Running a very large model requires expensive accelerators, power, engineering expertise, cybersecurity controls and continuous evaluation. For smaller organisations, a proprietary API may remain cheaper and simpler.
Open models also transfer responsibility. A customer using a closed service can rely partly on the provider for monitoring, filtering and security updates. An institution running an open model must establish its own safeguards, access controls, evaluation systems and incident-response procedures.
That is why the future is unlikely to be a simple victory of open AI over proprietary AI. Closed models will continue to lead in some areas, especially where enormous computing budgets and rapid research cycles matter. Their managed services will remain attractive to organisations that value ease of use over complete control.
But the strategic centre of gravity is shifting.
The Claude Fable shutdown demonstrated that intelligence delivered exclusively through a proprietary service can be withdrawn for reasons far beyond the customer’s control. GLM-5.2 demonstrates that increasingly capable alternatives can be downloaded, modified and operated without the same dependency.
The coming AI market may therefore resemble the history of computing itself. Proprietary systems will remain powerful and profitable, but open technology will become the foundation on which governments, companies and developers build systems they cannot afford to lose access to.
The decisive advantage of open models may not be that they always think better. It may be that their users can keep them.
