OpenAI cut the API price of its two lower-cost GPT-5.6 models on 30 July, dropping the entry-level Luna tier by 80% — from $1/$6 to $0.20/$1.20 per million input and output tokens — and the mid-tier Terra model by 20%. The flagship Sol model held its price but gained a “Fast mode” designed to process priority workloads up to 2.5 times quicker. What makes this notable isn’t the discount itself, it’s the timing: GPT-5.6 only launched on 9 July. Three weeks from release to an 80% price cut on the cheapest tier is a sign of just how much competitive pressure is pushing frontier AI costs down, and how quickly.

For SME owners, this isn’t really an OpenAI story. It’s a preview of what’s happening across the whole market, as rival labs undercut each other on price and pass efficiency gains straight through. If you’ve been quoted a cost for an AI-powered feature or tool six months ago, there’s a good chance the underlying model cost behind it has already dropped several times since.

Why “wait and see” has a real cost too

It’s tempting to treat rapid price drops as a reason to hold off — why commit to a build now when it’ll be cheaper in three months? But that logic cuts both ways. Every month spent waiting is a month a competitor could be spending building genuine advantage with a tool that’s already affordable enough to justify. The businesses winning here aren’t the ones timing the market perfectly on cost; they’re the ones building something useful now, on infrastructure and pricing that only gets more favourable to them over time.

There’s also a practical reason this matters more than it might first appear. A lot of SME hesitation around AI adoption over the past two years has come down to a genuine, reasonable worry about ongoing running costs — what happens if usage scales up and the monthly bill scales with it. That worry is becoming less justified with every price cut like this one. A task that looked marginal on cost six months ago, where the AI spend didn’t obviously clear the value it added, is worth re-running the numbers on now. The unit economics of AI-assisted work have shifted meaningfully since most SMEs last did that calculation properly.

What to actually check if you’re already paying for AI tools

If your business uses AI features baked into software you already pay for — a CRM, an accounting package, a helpdesk tool — ask your provider directly whether falling model costs are being passed on, or just absorbed into margin. It’s a reasonable question and one most SMEs never think to ask. If you’re commissioning something bespoke, whether that’s an internal tool, a customer-facing agent, or an automation that touches your core systems, get a clear view of how much of the ongoing cost is model usage versus development and maintenance — because the model usage line is the one that keeps shrinking. BuildApps works with UK SMEs on exactly this kind of build, and part of doing it properly is architecting around a model layer that’s actively getting cheaper, rather than locking in assumptions from six months ago.

It’s also worth remembering that price is only half the picture. A cheaper model is only good value if it’s actually doing the job well — the temptation with a steep discount is to reach for the cheapest tier by default, when the right call is usually to match the model to the task and let cost follow, not lead. Simple, high-volume work like drafting or summarising can comfortably move to a cheaper tier; anything touching judgement calls, customer-facing accuracy, or compliance-sensitive output is worth keeping on a stronger model even at a higher token price, because the cost of a wrong answer usually dwarfs the savings from a cheaper one.

The takeaway

Frontier AI pricing is falling fast and repeatedly, not as a one-off promotion but as a pattern. That makes this a good moment to revisit any AI project you shelved on cost grounds a few months back — the maths may already look different, and it’s likely to keep moving in your favour rather than against it.