Proprietary AI models' shelf life is too short
Most proprietary AI models from frontier labs seem to have a shelf life of ~12 months or less. That is an important issue for serious use, especially since AI models aren’t traditional software in the sense that new versions don’t have a defined scope of changes. A new model is an “upgrade” because it often performs better; however, it is a very different program. Thankfully, the API usually isn’t significantly modified, but the actual model is a new beast that requires engineer hours to carefully test and tweak the new version for your specific use case simply to achieve continuity. There are already many AI use cases that don’t really require higher intelligence. And what if you find out that “upgrading” to a newer version of an AI model actually degrades performance for your use case (while also coming with a higher price to make it feel extra frustrating)? This is entirely possible, especially if the lab in question started optimizing for something you don’t care about.
Google just announced it will be retiring the Gemini 2.5 range of models in October, including 2.5 Flash Lite, which is only a year old. With its low price, low latency, and acceptable performance, the only real competition comes from Chinese models (or perhaps second-tier labs now), and if you have any reservations about them, it was a natural choice. I myself use this model to power some functionality in my app for Polish firefighters. Google recommends Gemini 3.1 Flash-Lite, which is roughly 3× as expensive, as a replacement.
This is perhaps where open-weight and self-hosted models could shine, especially for institutions like banks and enterprises that value stability and predictability above new features, but I think it could also be beneficial to consider the labs’ perspective. (I’m speculating here; I don’t actually know what their internal rationale is.) Curtailing the use of older models makes your product range more manageable, especially if you’re releasing new ones every couple of months. Older models may also be more prone to making errors and thus be a liability. Perhaps a narrower range of models also lets you do some kind of infra wizardry by batching user requests, etc., and thus lowering inference costs.
Still, long-term support is an area where the frontier labs could and should improve, in my opinion, and I have two simple ideas for how to do it.
- Designate specific models as eligible for LTS, with upgrades required only every x years.
- Release your models to the public as open-weight models after retirement (or perhaps after x years if you’re worried about others catching up using your proprietary technology).