If you priced up AI tools for your business earlier this year and decided they were too expensive, it’s worth checking again. Over the past few weeks the three biggest AI labs have all moved on price in the same direction: down, or at least not up. OpenAI cut the cost of running its GPT-5.6 “Luna” model by up to 80% on input tokens. Anthropic scrapped a planned price rise for Claude’s Sonnet 5 model rather than pushing it through as scheduled. And xAI launched Grok 4.6 undercutting both rivals on cost per token. None of this made front-page news, but together it’s a real shift in what AI actually costs to run.

For SME owners this matters for a boring but important reason: most of the AI tools you touch, whether that’s a chatbot bolted onto your website, an AI feature inside your CRM, or a coding assistant your developer uses, are themselves built on top of these models. When the underlying model gets cheaper, the tools built on it either get cheaper too, get better for the same price, or the vendor pockets the difference. Knowing which one is happening to your subscriptions is worth five minutes of your time.

Why the price war is happening now

The short version is competition finally caught up with hype. A year ago there were really only two credible options for serious business AI use. Now there are at least four labs shipping frontier-quality models within months of each other, and none of them can afford to be meaningfully more expensive than the others for comparable performance. That’s normal market dynamics doing what they do, and it’s good news for buyers, even if the underlying technology story is more complicated.

It’s also a sign the industry is maturing past the “charge whatever the market will bear” phase. Anthropic’s decision to walk back a planned price rise rather than push it through is particularly telling, it suggests the labs are now competing on retaining business customers who have other options, not just on winning headlines about raw model capability. For a business owner, a market behaving that way is a healthier one to be buying from than one where a single vendor can set the terms unchallenged.

What this actually means for your business

Re-run the numbers on anything you decided against for cost reasons. If you looked at an AI-powered support tool, a coding assistant, or a document-drafting tool six months ago and the per-seat or per-token pricing didn’t stack up, it’s worth another look now rather than assuming the maths hasn’t moved. BuildApps works with UK SMEs specifically on matching the right AI tool to the actual problem, rather than buying whatever has the loudest marketing this month.

Don’t assume “AI” pricing is stable enough to build a five-year plan around. If a supplier’s roadmap depends heavily on a specific AI vendor’s current pricing, build in some slack. Prices have moved twice in two months this summer alone, in both directions depending on the vendor, and there’s no reason to expect that to settle down soon.

Cheaper doesn’t mean better for your specific task. The models getting cheaper aren’t necessarily the best fit for what you need. A cut-price frontier model is still worth testing against a smaller, purpose-built tool for jobs like invoice processing or customer email triage, where accuracy and integration matter more than raw model capability.

Check what you’re actually being charged for, not just the headline price. Some tools bill per seat regardless of usage, others pass the underlying token cost through directly. A cheaper model doesn’t automatically mean a cheaper bill if your vendor’s pricing hasn’t moved with it, so it’s worth asking the direct question rather than assuming the saving has been passed on.

The takeaway

The AI price war is a genuinely useful moment for SMEs sitting on the fence, not because any single price cut is huge, but because it confirms the market is competitive enough that vendors can’t simply charge whatever they like. If cost was the reason you shelved an AI project this year, it’s worth revisiting the numbers before you assume the answer is still no.