Why The Right AI Model Choice Cuts Your Costs

OpenAI has published a case study showing legal tech firm LegalOn cut its daily Codex costs by 65% without slowing down its development work. Codex is OpenAI's AI coding assistant, available in different model tiers: Astra (the most capable, most expensive), Sol (mid-tier) and Luna (fastest, cheapest). LegalOn's engineering team matched each coding task to the model that could actually do the job, rather than defaulting to the top-tier model for everything. The 65% figure is OpenAI's own number from the case study, and it held while LegalOn kept the same development speed, so this was not a trade-off of speed for cost.

What this means for your margin

If you run a software team, or you pay an outside developer who bills you for AI-assisted coding time, this matters directly to your invoice. Most teams default to the most powerful model because it is the safe choice. That is like hiring a structural engineer to check a fence post. It works, but you are paying senior rates for junior work.

Worked example: say your developer runs 200 Codex tasks a month and every one goes through the top-tier model at an average cost of 10p per task, that is £20 a month on a tiny project, or £2,000 a month on a team running 20,000 tasks. Route the simple 70% of those tasks, bug fixes, formatting, test writing, to a cheaper model and you could be looking at a bill closer to £700, roughly LegalOn's reported 65% reduction, for the same output.

This is not just a software story. Daniel Walsh owns a software firm alongside his parts and construction businesses, and the same principle applies wherever you pay by AI usage, whether that is coding, document review or quote generation. The lesson is simple: the model you default to is a cost decision, not just a technical one, and nobody checks it unless you ask.

What to do this month

• Ask your developer or software supplier which Codex model they default to, and whether they route tasks by complexity.
• If you commission AI-assisted development work, ask for a breakdown of model usage on your invoice.
• If you run your own Codex account, test routing simple tasks to a cheaper model tier and track the cost difference over a month.