The Price of AI: The Discount You Pay With Your Data

In three days, three labs released a frontier model: GPT-6 Astra, Claude Fable 5.1, and Meta's Muse Spark 1.3. What really changed wasn't the intelligence of any of them. It was where the bill gets charged.

Until now, comparing models meant comparing two lines: price per million input tokens and per million output tokens. Those two lines no longer tell you how much you'll pay.

What each one charges today

• GPT-6 Astra (OpenAI): US$10 per million input tokens and US$50 for output, with cached input at US$1.
• Claude Fable 5.1 (Anthropic): the same US$10 and US$50 on the price list. What changed was caching, which dropped from US$1 to US$0.25 per million, a 75% cut.
• Muse Spark 1.3 (Meta): US$1.25 for input and US$4.25 for output, a fraction of the other two. And there's a second, much lower price, explained below.

Looking at the price list alone, the conclusion would be simple: Meta is ten times cheaper. The real conclusion depends on two things the price list doesn't show.

Caching got 75% cheaper, but only for those who built it right

Caching, here, is the chunk of text you send repeated in every call: the system prompt, the product manual, the catalog, the conversation history. The model stores that processed chunk and charges much less to read it again.

With Anthropic's cut, reading from cache now costs 2.5% of the standard input price. The practical effect reported is a drop of around 25% in the cost of a typical workload, and up to 45% in heavily agentic tasks — the ones where the model makes many calls in a row, rereading the same context at every step.

The detail that decides whether you actually get this discount: the cache only kicks in when the beginning of what you send is identical from one call to the next. If your system assembles the text with the date and time at the top, or shuffles the order of the instructions, or injects the customer's name before the manual, the chunk changes with every call and there's no caching at all. You pay full price and never find out why.

In other words: this price cut isn't a discount that automatically lands on your bill. It's a discount for whoever orders their text a specific way — fixed content first, variable content at the end.

The over-90% discount you pay for with data

Meta made the most aggressive move, and the most interesting one to analyze. Beyond the standard price, it created a tier for those who authorize the company to use your prompts and the model's responses to train its future models.

In this tier, a million input tokens costs US$0.10 and output costs US$0.20. Compared to Meta's own standard price, that's a drop of around 92% on input and 95% on output. Against the competitors' US$10 and US$50, the gap exceeds a hundredfold.

This isn't charity or price war in the usual sense. It's Meta putting a price on your data and letting you decide whether you want to sell it. The company also signaled it plans to release the model's weights — which is a promise, not a done deal, and should be treated as such.

Why this discount may be off-limits for you

Here the decision moves from finance to legal. If what you send the model includes customer data, taking the discount means authorizing a third party to train on that data. Three questions settle the matter for most companies:

1. Do you have a legal basis for it? Brazil's LGPD requires one. Did the customer consent to having their data processed to train another company's model? The answer is almost always no.
2. Does your contract allow it? Contracts with corporate clients usually include a confidentiality clause that already rules this out, without even reaching the LGPD question.
3. What exactly goes along with it? It's not just the prompt. It's the document you attached, the system snippet the agent read, the name that showed up in the middle of the conversation.

There is a use case where the cheap tier is perfectly fine: content that's already public or that doesn't identify anyone — generating a catalog product description, summarizing text you published yourself, testing a prompt on fictional data. The mistake is defaulting to the cheap tier and finding out afterward what went through it.

How to choose now

With three variables in play — list price, caching, and data — comparing by token price has become useless. What actually works:

1. Measure cost per completed task, on your task, with your text. It's the only measure that combines all three variables.
2. Split your workloads in two. The ones that pass sensitive data and the ones that don't. They can, and probably should, run on different models.
3. Before switching providers over caching, look at your code. If the start of the text changes with every call, the discount doesn't exist for you — and reordering your text is cheaper than switching models.
4. Treat the open-weights promise as a promise. Planning a migration around something not yet released is like signing a contract with the date left blank.

This is the kind of analysis we do in AI implementation for businesses: measuring your task before choosing the model, and separating what can and can't leave the building. For more on the OpenAI launch, see what really changes with GPT-6 Astra; on the cost of running an open model on your own machine, how much it costs to run an open model. And the framework for what to let AI access is in the permission checklist.

Sources

Announcement of Claude Fable 5.1 and Mythos 5.1 and Anthropic's pricing documentation; the GPT-6 Astra page and OpenAI's model documentation; Meta's Muse Spark 1.3 pricing materials and coverage from VentureBeat, DataCamp, and MindStudio on both pricing changes. The tier discount figures were checked outside the newsletter that originated this story.