Open models win on tokens, closed models win on dollars

Hope everyone is having a great weekend!

If AI is still on your mind and you’re looking for some food for thought, I was looking at open vs. closed-weight models on OpenRouter this morning, and the numbers are thought-provoking:

• Traffic: Since last October, closed-weight model traffic has grown 11.6x, while open-weight traffic has grown 77x.
• Token share: Over the same period, closed-weight models’ share of tokens dropped from 72% to 28%.
• Spending: Despite that shift, closed-weight models still captured 73% of total spending in September.

In other words, open models are winning on tokens, while closed models are still winning on dollars.

With Opus 5.5 being the most popular model amongst my enterprise friends, and Haiku 5.5 now costing only one-tenth as much as Haiku 4.5, I have no idea which way the charts will move over the next 12 months.

One metric I’d really like to see is the cost per completed task, including the retries and reviews.

Has anyone come across data that measures it on real workload across open and closed-weight models?

Enjoy the rest of your weekend! Chart below.