Dario Amodei's strategy, summed up in the provocative formula 'too powerful for you users', has not held up under scrutiny. The post reporting the result is blunt: two open models lead the Artificial Analysis Cyber Index and beat every single model from Anthropic. GLM 5.3 Flash sits at the top; Mistral Large 4, another open model, also ranks better than its closed rivals.

The result is more than a leaderboard story. When an open-weights model reaches the top of an independent index, the incentives shift for anyone choosing a platform. Companies that currently pay for closed-model APIs must ask whether the performance gap still justifies dependence on a single provider. For many deployments, especially on-premise, open weights offer a structural advantage: control over the model, the ability to run inference locally without sending data to third parties, and a variable cost that does not scale with every token.

This is not a minor detail. The main constraint for open-source models is often no longer absolute quality, but hardware management: you need GPUs with adequate VRAM, an efficient serving pipeline, and a quantization strategy to reduce the footprint without losing too much. These are known problems, but the news changes the TCO calculation: if an open model delivers performance comparable to or better than a closed one, the marginal cost of inference moves from API fees to infrastructure, which many organizations have already amortized or can share.

The overtaking also has a second-order effect on competition. Anthropic has built its communication around cautious releases, but a leaderboard dominated by open models risks pushing developers and system integrators toward self-hosted solutions. For cloud vendors that resell closed models, the pressure is twofold: they must justify the premium and compete with open ecosystems that improve quickly. It is no coincidence that the post says 'Open source prevails': it is not a neutral statement, but a symptom of shifting preferences.

Structurally, the result signals that the AI frontier is no longer the monopoly of a few labs with unlimited budgets. The availability of competitive open models reduces lock-in and increases data sovereignty: a company can train or fine-tune on an open base model, keeping sensitive data within its own perimeter and adapting behavior to its needs. For those evaluating on-premise deployment, this does not remove orchestration complexity, but it shifts the decision from 'which API to buy' to 'which stack to build'. AI-RADAR offers frameworks on /llm-onpremise to weigh the trade-offs.

It remains to be seen whether the leadership of GLM 5.3 Flash and Mistral Large 4 on the index will translate into real enterprise adoption. The index measures one dimension, but moving to production requires robustness, controlled latency, and a mature tooling ecosystem. The direction, however, seems clear: open models are no longer the fallback option, but the starting point for those who want control and independence.