AI Got 40% Cheaper to Run. Your Quote Is Out of Date.

On 22 September, Anthropic cut the price of its most capable model by 20% and said it runs about 40% cheaper on typical work. If someone quoted you for an AI assistant in the last year, part of that quote is now out of date.

Not all of it. That distinction is the whole point of this post, and it is the part you will not get from a vendor.

What actually changed

Claude Opus 5.5 landed at $4 per million input tokens and $20 per million output, against $5 and $25 before. Anthropic says it performs at the level of their previous top model while costing around 40% less to run, with output roughly 30% faster and cached reads down 60%. Sonnet and Haiku versions are expected within weeks, and those are the cheaper tiers most business systems actually run on.

This is not a one-off. Model prices have fallen repeatedly for two years while capability has gone up. What is different here is the size of the step and that it landed on the tier businesses use.

What this does not change

The build cost. On most AI projects we quote, the model is a small fraction of the total. The bulk is working out what should happen, connecting it to the tools you already run, writing the rules that stop it saying something wrong, and testing it against real cases. None of that got cheaper yesterday.

So if you were quoted £3,000 to build something, a 40% drop in model pricing does not make it £1,800. Anyone telling you otherwise is either confused or selling.

What it changes is the running cost — the monthly bill for the thing once it exists. That is the number that quietly compounds, and it is the number that just fell.

When it is worth revisiting a system

Three situations where the maths genuinely moved:

Your monthly AI bill is a real line item. If you are spending enough on model usage to notice it, a 40% reduction is worth an hour of someone’s time. If it is a few pounds a month, leave it alone.

You were told something was too expensive. Plenty of businesses got told in the last year that a particular feature did not justify its running cost. Some of those are now fine. Voice, longer documents and anything processing high volume are the usual candidates.

You are on an older model for cost reasons. If someone put you on a weaker model purely to keep the bill down, the better one may now cost what the weaker one used to. That is a straight quality upgrade for the same money.

And the situation where it is not worth it: your system works, the bill is small, nobody is complaining. Changing a working AI system carries real risk of it behaving differently. A saving of £15 a month does not justify that.

The part most people get wrong

A price cut is not a reason to start using AI.

If a system did not make sense at the old price, a cheaper model rarely fixes that, because the reason was almost never the model bill. It was that the task was not well defined, or the data was not there, or nobody had time to check the output.

Cheaper models make a good case better. They do not rescue a weak one.

What we would tell you on a call

If you are running something we or anyone else built in the last year, we will look at what it costs you now against what it would cost on current pricing, and tell you whether the difference justifies touching it. Half an hour, no charge. Most of the time the honest answer is leave it, and we will say so.

If you have been putting off building something because the running cost did not work, that is worth a second look now. The number moved.

Related reading: You’re being moved to an AI plan you didn’t ask for covers the opposite problem, where the price went up without you choosing it. How much AI automation actually costs is the honest version of what building this runs to. And why your software bill went up in 2026 puts it in context against everything else that got more expensive this year.