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 makes the trade-off between price and data privacy explicit rather than implicit. 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, those same tokens cost 10 cents and 20 cents respectively -- roughly a 95% discount -- in exchange for every prompt and model output a developer sends being used to train Meta's future models. Cached-token pricing drops even further, from $0.15 to $0.002 per million, a figure Cryptopolitan's Randa Moses independently confirmed at a "75x reduction."

The catch that isn't in the discount

The price cut isn't the only difference between the two tiers. Contributor-tier access caps requests at 100 per minute, against 3,000 per minute on the standard plan -- a 30x throughput cut that Moses notes limits the tier to prototyping and low-volume experimentation rather than production workloads. That framing lines up with how Meta's own pricing documentation describes the tier, quoted by Fernholz: it "lowers the barrier to entry for prototyping, testing integrations, and scaling experiments."

Why Meta needs this data specifically

Meta has struggled to source the kind of real coding-agent interaction data that improves agentic performance. Both outlets note the same recent context: an internal initiative to track employees' own computer usage for training data, launched earlier in 2026, drew enough internal criticism that Meta paused it in June. The Contributor tier is a different approach to the same underlying problem -- instead of monitoring people who didn't opt in, it prices the opt-in explicitly and lets developers choose it workload by workload. Meta declined to comment on the new pricing model when TechCrunch asked.

The bill this doesn't show up on

Every major lab already trains on some tier of user data by default and treats not doing so as the premium option a privacy-conscious customer pays for. Meta's Contributor tier inverts that: opting into training is now the discount, not the default, and the discount is large enough -- 95% on tokens most developers already pay for constantly -- that the actual cost of choosing privacy becomes visible on a rate card instead of buried in a policy document nobody reads before clicking accept. For any organization tracking what its AI spend actually buys, that's a real trade-off to price deliberately, not a checkbox to leave on the default setting.