A write-up doing the rounds on Hacker News this weekend lists the price changes from OpenAI and Anthropic between June and the end of September, and it's quite a list. So it seemed a good moment to explain what's changed and what to check.

What's actually happened to prices

I've checked the main figures against the vendors' own pages, because I don't like quoting prices second hand. All prices are in US dollars per million tokens (a token is roughly three quarters of a word).

  • Anthropic's Claude Opus 5 launched in July at $5 for input and $25 for output. Claude Opus 5.5 followed on 22 September at $4 and $20. Anthropic says that on typical workloads it "will cost 40% less than Opus 5" at default settings.
  • OpenAI's GPT-5.6 Sol is listed at $4 for input and $20 for output. GPT-6 Sol, which arrived in September, is listed at $2 and $10. That's half the price.
  • OpenAI's smaller models have dropped too. GPT-5.6 Luna is listed at $0.20 for input, and GPT-6 Luna at $0.10.
  • Cached input has fallen as well. GPT-6.1 Sol charges $0.10 for cached input against $0.20 for GPT-6 Sol, and Opus 5.5 charges $0.20 for cache reads.

Why it's happening is anyone's guess, and the write-up admits it can't fully explain it either. Competition is fierce and it's easy to swap one model for another, so that probably plays a part. For a small business, though, the why matters a lot less than the what. The same work costs a good deal less than it did in June, if you're on the right model.

Why your bill probably hasn't moved

When someone builds an AI feature, they pick a model and put its name in the code or the settings. That's the right thing to do. You want the model fixed, because a different model can behave differently, and you don't want your app's answers changing overnight without anyone deciding they should.

The side effect is that the app keeps using that exact model until a person changes it. In most of the examples above, the cheaper price came with a new model, not a cut to the old one. Opus 5 is still $5 and $25. GPT-5.6 Sol is still $4 and $20. If your app was built in the summer and nobody has touched it since, it may well be paying summer prices.

That's not anybody doing anything wrong. It's just that "check whether there's a cheaper model" doesn't sit on anyone's to-do list unless you put it there.

What to check

None of this needs a big project. It's mostly a short list of questions for whoever looks after your app.

  • Which model, exactly, are we using? Not "we use Claude" or "we use OpenAI". The specific model name. Then compare it with the provider's current pricing page.
  • What did we spend last month, and on what? If you can't break the bill down by feature, that's worth fixing first. We wrote about that in how to stop your AI bill running away.
  • Is caching switched on? If your app sends the same long instructions or documents with every request, caching can make the repeated part far cheaper. On the models above, cached input costs a tenth of the normal input price or less. It often isn't turned on in a first version.
  • Does every job need the big model? Sorting, tagging and summarising short bits of text can often be done well by the smaller, cheaper models. Save the expensive one for the work that needs it.
  • How hard would it be to change? If the model is a single setting, trying a newer one is a small job. If it's scattered through the code, that's a problem worth knowing about. There's more on that in our post on keeping your AI provider swappable.

Don't just flip the switch

Here's the bit people skip. A newer, cheaper model isn't automatically a drop-in replacement. It might phrase things differently, follow your instructions a little differently, or give longer answers. Longer answers mean more output tokens, and output is the expensive part, so a lower price per token doesn't guarantee a lower bill.

The sensible way to do it is boring but reliable. Take a decent sample of real requests your app handles, run them through the old model and the new one, and compare both the answers and the cost. If the new one does the job as well and the sums work out, switch. If not, you've lost an afternoon and learnt something.

And keep an eye on retirement dates. Providers do eventually stop offering older models, and it's much nicer to move on your own schedule than on theirs.

Where we come in

If you've got an app with AI in it, perhaps built quickly with one of the AI tools, our AI Code Audit will tell you how it's put together, including which models it calls, how the AI costs stack up and how easy it would be to change. It's £250 plus VAT for one system, and you get a written report within five working days with a fixed price for each fix.

If you'd rather somebody just kept an eye on this sort of thing for you, along with security updates, backups and the rest, that's what our support and maintenance plans are for.

The short version

AI has got cheaper, but the savings don't turn up on their own. Find out which model your app uses, check what the newer ones cost, test before you switch, and make sure caching is on. If you'd like a hand with any of it, give us a ring on 01474 639 089 or drop us a line.

Cheers, Rob