Meta will cut your AI bill 95% if you let it train on everything you send

Meta has introduced a two-tier pricing structure for Muse Spark, its coding and agent model, that puts a price on the trade between cost and data privacy. TechCrunch's Tim Fernholz reported on September 3, 2026 that Meta's standard rate is $1.25 per million input tokens and $4.25 per million output tokens. Under the new "Contributor" tier, the same tokens cost 10 cents and 20 cents respectively, roughly a 95% discount, and every prompt and model output a developer sends is used to train Meta's future models. Cached-token pricing drops from $0.15 to $0.002 per million, a figure Cryptopolitan's Randa Moses independently confirmed as a "75x reduction."

The rate-limit catch

Contributor-tier access caps requests at 100 per minute, against 3,000 per minute on the standard plan. Moses notes that 30x throughput cut limits the tier to prototyping and low-volume experimentation. Meta's own pricing documentation, quoted by Fernholz, describes the tier the same way: it "lowers the barrier to entry for prototyping, testing integrations, and scaling experiments."

Why Meta wants this data

Meta has struggled to source the coding-agent interaction data that improves agentic performance. Both outlets note the recent context: an internal initiative to track employees' computer usage for training data, launched earlier in 2026, drew enough internal criticism that Meta paused it in June. The Contributor tier attacks the same problem from the other side. It prices the opt-in and lets developers choose it workload by workload. Meta declined to comment on the new pricing when TechCrunch asked.

Privacy now has a line-item price

Every major lab already trains on some tier of user data by default and sells not doing so as the premium option. Meta's Contributor tier flips that: opting into training is the discount. At 95% off tokens most developers pay for constantly, the cost of choosing privacy shows up on a rate card instead of in a policy document nobody reads. If you track what your AI spend buys, decide this per workload. Don't leave it on whatever the default is.