The cheap tier on Muse Spark is paid for with the developer's data

One model, two prices

On the contributor tier Meta opened for Muse Spark, 1 million input tokens cost 0.10 dollars and 1 million output tokens cost 0.20 dollars. The same amounts cost 1.25 dollars and 4.25 dollars on the standard tier. What closes the gap is no hardware gain and no smaller model: it is the developer allowing their prompts and the model's outputs to be used in reinforcement learning. The token price turns into a choice of payment method — cash or data.^[1]

That creates a sorting job a developer has to do before the first request. Which prompt is shareable, which one carries customer data, which output reveals behaviour of a product that has not been announced? Meta's pricing guide frames the tier for cases where training on your data is acceptable, and that framing hands the decision to the developer. Arvind Narayanan's observation points the same way: large companies tend to pick the expensive plan, citing data retention and enterprise IT governance. The real buyer of the cheap tier may be the small team with no legal or security staff to do that sorting — an inference, and the opposite is possible. Because the sorting is tedious, small teams may instead move to the expensive tier and buy the problem away.^[1]

The second line, due in October

On October 2, 2026 GitHub removes Gemini 3.5 Flash, Gemini 3.6 Flash, Kimi K2.7 Code and Claude Opus 4.7 from every Copilot surface and points users to Gemini 3.8 Flash, Kimi K3 and Claude Opus 5. The scope is wide: Copilot Chat, inline edits, ask and agent modes, code completions. Copilot Enterprise and Copilot Business administrators may need to enable access to the alternatives separately through model policies, so the timetable brings configuration work as well as a version change.^[2]

The constraint the two developments share is that the falling cost of the model layer is tied to two needs on the vendor's side: training data and model lifecycle. Meta lowers the token price and asks for the developer's stream of prompts and outputs in return; GitHub leaves list prices alone but decides which models stay on the list and when they go. In both cases the price the developer pays does not appear on the pricing page: one is a data classification and governance decision, the other is moving prompts, evaluations and agent scaffolding onto a new model. On a bottleneck map that is a constraint leaving model selection and settling into governance and migration work.^{[1], [2]}

What stays in the builder's hands

The practical consequence is plain: when a team compares token costs it now has to open two more columns. The first is who decides which prompts are shareable and how that decision is audited. The second is how many days it takes to move prompts and evaluation sets when the model in use leaves the list. A cost calculation without those two columns makes a 95 per cent discount look cheaper than it is.^{[1], [2]}

There is an observable signal too. If Meta publishes the share of users on the contributor tier, or the token volume coming from it, by the end of the year, whether the cheap price actually brings in new developers becomes assessable from outside. Until such a number is published, who the discount attracts stays information held on the vendor's side. What is available today is the price table and the condition printed under it.^[1]