I do not use Claude or OpenAI day to day. The reason is arithmetic, not prejudice.
My heavy load runs on open models. GLM on my infrastructure and on Z.AI’s plan, for a fraction of what a closed subscription would cost. Cost, quota, and multimodal solved in the same place.
The math the closed models lost
The closed models’ argument was capability: the better model justified the price. That argument died when the open model tied on the task and won on the cost structure:
- Quota and window: the open plan carries high quota per window with half-price off-peak, which changes the invoice for heavy users.
- Multimodal included: text, image, video, and file in the same subscription, without a separate product per modality.
- The escape hatch: published weights mean the subscription has a plan B. A closed key is a key only the vendor turns off.
For my volume (agents running daily, code, documents), the sum of these layers makes the difference the per-million price hides. The ruler is cost per completed task, and per task the open model won.
The per-use exception: architecture, not dogma
The open model does not win everything, and honest design admits it. When a task requires what only the closed model has (a proprietary API, a specific behavior, a stamp a client demands), I pay that task per use and return to open.
It is an exception with reason, not a habit. And it is exactly the inverse of the common pattern: paying an expensive closed subscription every month to use a fraction the open model would cover. A model’s commercial life teaches the arc; the open-versus-closed math is where the arc becomes a decision.
The test everyone should run on their own load
The final answer is not mine, it is each user’s measurement on their task:
- Inventory the real weekly load (code volume, documents, agents).
- Test the open model for two weeks with per-completed-task measurement.
- Add the layers (quota, off-peak, multimodal, self-hosting) on both sides.
- Define the boundary: what stays open, what becomes a per-use exception, and why.
If you live on Claude or GPT and the math closes for your volume, perfect. For my volume, closed became a niche. Whoever wants to run the same math on the plan I use: https://z.ai/subscribe?ic=BGIRP61T2P
Conclusion
The era when closed won by default is over: the open model tied on the task and won on structure. The surviving decision is not faith, it is a spreadsheet.
My design is open by default and closed as a paid-per-use exception. Yours can be different, as long as it comes from measuring your own load and not from subscription habit. October’s question is the same since September: how much does your completed task cost, and under which ruler?