What happened
Legal tech company LegalOn has cut its spend on OpenAI's Codex API by 65%. The company achieved this without slowing down its development cycle. According to a case study published by OpenAI, the savings came from a simple strategy — matching the right model to the right task. Instead of using the most powerful model for every job, their team routed tasks to different Codex models based on complexity.
How the room's reading it
Builders are reading this as a practical guide to cost control. On platforms like X, developers are pointing to it as a clear example of moving beyond the 'use GPT-4 for everything' mindset. The consensus is that tiered model usage is smart, but often hard to implement well — so a concrete success story is useful. There's some scepticism, of course. Cynics note that the case study comes directly from OpenAI, framing their own product tiers as the solution. Still, for teams feeling the pressure of rising API bills, the post is being shared as a welcome, practical playbook.
Sailfish's take
We've seen this pattern with teams we work with. It's tempting to throw the most powerful model at every problem, but it's almost always the wrong long-term move. The real engineering challenge isn't the prompt — it's building the evaluation and routing layer that sits in front of the models. LegalOn's story validates this approach. This isn't just about saving money, it's about building more resilient systems. Smaller models are often faster and more reliable for simpler tasks. If you're building a new AI feature, don't start with the biggest model and work down. Start with the smallest, cheapest model that gets the job done.