Meta Just Put a Price on Your Prompts, and It Is Not 95%

Key takeaways

• Standard Muse Spark pricing is $1.25 per million input tokens and $4.25 per million output tokens.
• Contributor pricing, where you share prompts and outputs for future model development, is $0.10 per million input and $0.20 per million output.
• That is 92% off input and 95.3% off output. The headline number quoted is about 95% on average, but almost nobody sits exactly on the average.
• Output-heavy agent workloads capture more of the discount than context-heavy retrieval workloads.
• The real decision is not the discount, it is whether your prompts are an asset you are giving away.

Everyone quoted the 95% and moved on

Most AI tools let you opt out of sharing your usage. Meta has taken that switch and attached a price to it. For Muse Spark, its new model aimed at running coding and other agents, there is an explicit contributor tier: cheaper tokens in exchange for your prompts and outputs feeding future model development.

The reported average works out to roughly a 95% discount. Fine. But an average is not a bill, and the two token types are not discounted equally.

Input goes from $1.25 to $0.10 per million. That is 92% off.

Output goes from $4.25 to $0.20 per million. That is 95.3% off.

That gap is small in percentage points and large in consequence, because your blended discount is a weighted average of the two, weighted by your own traffic mix. Nobody in that deal is getting exactly 95%.

Who actually captures the bigger discount?

Output-heavy workloads. Agents that generate long plans, write code, draft documents, or loop through many generation steps skew toward output tokens, and output is the more deeply discounted line.

Context-heavy workloads capture less. A retrieval-augmented app that stuffs 100,000 tokens of context in to get 2,000 tokens back is buying mostly input, and input is the 92% line.

An illustrative pass, so the shape is visible. Say a request uses 100,000 input tokens and 20,000 output tokens.

• Standard: $0.125 of input plus $0.085 of output, so $0.21 per request.
• Contributor: $0.010 of input plus $0.004 of output, so $0.014 per request.
• Effective discount: about 93.3%.

Swing the mix toward generation and you drift up toward 95%. Swing it toward long context and you drift down toward 92%. The point is not that the difference is enormous, it is that you should be computing your number rather than quoting Meta's.

Why is Meta paying for this at all?

Because agent training data is scarce and expensive to obtain. Meta's earlier attempt to gather usage data internally, by tracking employee computer usage, drew heavy internal criticism and was paused in June.

The TechCrunch piece quotes Mario Zechner, developer of the open source harness Pi, arguing that a big jump in coding agent capability came from an assistant storing agent sessions by default and using them for reinforcement learning. Whatever you make of that specific claim, the structural point holds: the models that see the most real agent traffic improve fastest, and model builders know it.

Meta's own pricing guide frames the contributor tier as lowering the barrier for prototyping, testing integrations and scaling experiments where training on your data is acceptable. That is a fair description. It is also an admission of what is being bought.

What is your prompt data actually worth?

Here is the uncomfortable part, and the reason this is a pricing story rather than a privacy story.

Princeton's Arvind Narayanan pointed out that large companies stick with token-billed enterprise plans even though consumer subscription plans are discounted by 10x to 20x or more, with the main differences being data retention and enterprise IT governance. Those companies are already paying an enormous implicit premium to keep their data out of training. They just never wrote the premium down as a line item.

Meta has now written it down for them. If your enterprise is paying 10x to avoid data sharing on one vendor while considering a 95% discount to accept it on another, you do not have a data policy. You have two contradictory ones.

The useful exercise is boring and takes an afternoon. Split your workloads into three buckets: prompts that contain genuinely proprietary content, prompts that contain customer data you are contractually obliged to protect, and prompts that are essentially generic scaffolding around public information. Most teams find the third bucket is much larger than they assumed. That bucket is where a contributor tier is close to free money, and the other two are where it is not worth 95% or anything like it.

Does this change the wider pricing picture?

It adds a second axis to price competition. Until now, frontier labs mostly competed on headline per-token price, and there has been plenty of movement there: Anthropic's newest Fable and Mythos models arrived with lower costs for processing cached tokens, and OpenAI's latest models took major price cuts at the end of July.

Data-for-discount is a different lever. It lets a provider quote a dramatically lower number without touching its standard rate card, and it segments customers by something other than volume. Expect imitators, and expect the discounts to be quoted as single averages that do not survive contact with a real traffic mix.

The takeaway: a headline discount is a marketing number, your blended rate is the real one, and the gap between them is where pricing decisions actually live.

If you want to see what a contributor tier would do to your blended rate at your own input to output split, you can run both price sets side by side in the Calcaas provider comparison.