LLM API pricing: why we publish dated prices

"AI is getting cheaper." You've probably read that sentence many times this year. Cheaper than what? Since when? For which model, and for whom?

A price without a date is a rumour. That's the idea behind Model Watch, the series on our Research page where we publish model prices with the day we read them. This piece explains why we do it that way, and how to use the numbers.

What Model Watch is

Model Watch is a running record of what AI models cost to use. Each post takes a set of models, reads their published prices from the provider's own pages and writes them down, with the date.

Our first post, Model Watch #1, looked at paid models. As of 22 October, four paid models list $2.00 per million input tokens, with outputs from $6 to $10.

That sentence has four parts, and each one is there for a reason.

Why every part of that sentence matters

"Four paid models." We say exactly what we looked at. Not "the market," not "leading models." A reader can see the sample and decide whether it's relevant to them.

"List." These are the prices each provider publishes. They are not discounted contract prices, and they are not our prices. Starfield doesn't sell these models, and nothing in Model Watch is a comparison with us.

"Per million input tokens" and "outputs." AI is usually priced by the token, roughly a word or part of a word, with different prices for what goes into the model and what comes out. A single headline price hides that difference. Writing both down lets a reader estimate the real cost of their own workload, which might be mostly reading, mostly writing or both.

"Read on 22 October." This is the part most price comparisons leave out, and it's the part that makes the rest useful.

Why the date is the point

AI prices move. Providers cut prices, launch new tiers, retire old models and change how they count. A price list that was accurate in one month can be wrong the next.

Without a date, there is no way to tell whether a number is current, out of date or simply wrong. With a date, three useful things become possible.

You can trust it for what it is. A dated price is a claim about one moment. It doesn't pretend to be permanent, so it can't quietly go stale.

You can compare over time. One reading is a snapshot. A series of readings, taken the same way, shows a direction. That is how you find out whether "AI is getting cheaper" is true, for which models and by how much.

You can check our work. If you visit a provider's page and see a different price, the date tells you whether it changed after we read it or whether we made a mistake. Either way, you know what to do with the number.

How we read a price

We keep the method simple so anyone can repeat it. The rules we publish by are set out in what we will and won't claim here.

1. Go to the source. We read prices from each provider's own published pricing page, not from third-party summaries.
2. Record the date. The day we read the page goes next to every figure.
3. Keep units consistent. We convert to the same unit, per million tokens, and keep input and output separate.
4. Say what's excluded. If a provider offers volume discounts, batch pricing or regional differences, we note what we did and didn't include.
5. Publish corrections. If we get a number wrong, we fix it and log the change, with the date.

The full method sits alongside each post, so the steps behind a figure are always one click away.

How to use the numbers

If you're budgeting for an AI feature, Model Watch can help in a few practical ways:

• Estimate a range. Take your expected input and output volumes, apply the listed prices and you have a starting figure for each model.
• Check a quote. If a vendor's offer seems high or low, compare it with the dated list prices. A big gap is a good reason to ask questions.
• Watch the trend. Before you commit to a long contract, look at how prices for that class of model have moved across several readings.
• Look past the price. A cheaper model that needs more tokens, or runs further from your users, may cost more in practice. Price is one input to the decision, alongside speed, location and where your data is allowed to go.

Why a company like ours publishes this

We build AI infrastructure and describe ourselves as operators, not developers, so it's fair to ask why we spend time recording other companies' prices.

The answer is simple. Buyers can't make good decisions without good information, and right now much of the information about AI costs is vague. We'd rather earn trust by publishing numbers people can check than by making claims they can't. When our own systems are running real workloads, their numbers will follow the same rules: method, date and source.

What we're building

Starfield is building standardized compute close to power, connectivity and demand, connected through one programmable network for AI inference. It is designed so every workload can generate an auditable record of commitment, placement and delivery.

Read how Model Watch gets its numbers, and follow along as the readings add up. To talk about your own workloads, use Request capacity below.