I'm writing a lot of software - or at least I'm guiding AI coding tools in writing a lot of software.
Part of this is interacting with different models. Since I'm using Anthropic's Claude I usually encounter Opus, Sonnet, and Haiku. I use all of these daily, as well as OpenAI ChatGPT (ongoing conversations) and GitHub Copilot for Pull Request reviews[1].
It's tempting to just reach for "the best AI", which you could argue is Opus for coding. But Software Engineers have realized for a while that this can be a bad call. Opus is great for planning and architectural reasoning. Sonnet is fast and "good enough" for iteration. Haiku handles lightweight tasks like commit messages and lookups.
In other domains, this kind of model selection is less established. If your AI policy says "we use X" or your procurement team has only approved one vendor, there may be good reasons for that — but you may also be locking into a suboptimal outcome. Worse, if your training and onboarding only covers one tool, you're missing out.
- Cost profiles vary wildly. A task that costs dollars on a frontier model might cost cents on a smaller one. Pick the wrong model and you pay too much or get too little.
- Different models have different strengths. Sometimes a model might simply be much faster, sometimes more accurate. Others thrive in a specific domain, others are generalists.
- Model diversity can be important. In my case, the use of Copilot not only provides great reviews, it's an explicit counterbalance to my overall heavy use of Claude.
- Context matters. Security-sensitive code might warrant a more capable (and auditable) model. Boilerplate generation probably doesn't.
Policies that ignore this get ignored, or slow people down.