Hisham AlhussainAugust had one of those weeks where you can barely keep up. OpenAI released gpt-oss, its first open-weight models in years, on the 5th. Anthropic shipped Claude Opus 4.1 the same day. Then GPT-5 arrived two days later.
Most of the attention went to GPT-5, which makes sense. But for a lot of the teams I work with, gpt-oss was the bigger news.
Why open weights matter so much here
In Saudi Arabia, and in a lot of regulated industries everywhere, data residency is a real constraint, not a checkbox. Some data simply can't leave certain environments. For those workloads, the conversation used to end quickly: "We'd love to use a frontier model, but we can't send this data out."
Strong open-weight models change that. You can run them in your own environment, on your own hardware or your own cloud region, and keep the data where it needs to be. They're not frontier-level at everything, but for a lot of real tasks like classification, extraction, and answering questions over internal documents, they're more than good enough.
So, build or host?
Here's roughly how I'm thinking about it now:
- Sensitive data, well-defined task: a self-hosted open-weight model is a serious option. Validate it with your own evals first.
- Hard reasoning, low sensitivity: use the best hosted model you can. GPT-5 and Opus 4.1 are genuinely better at the hard stuff.
- Mixed workloads: route. Keep the sensitive steps in-house and send the rest out. More on routing another time.
The thing I'd warn against is picking a side on principle. "We only use open models" and "we only use API models" are both ways of skipping the actual analysis.
Hosting isn't free
Self-hosting sounds cheaper until you price it out properly. GPUs, autoscaling, monitoring, security patching, and someone on call when it falls over at 2am. For low or bursty volume, a hosted API often wins on total cost even when the per-token price looks higher.
A word on routers
GPT-5's built-in router, which decides how much thinking to do for each request, is interesting too. It also means the same prompt might get handled differently from one day to the next. If you rely on consistent behavior, pin your settings and keep your evals running. Assume nothing stays still.
Honestly, this was the most exciting month in a while. Having real choice between hosted and self-hosted is a big deal for the kinds of projects I care about.