Meta charges 12.5 times more for Muse input tokens it promises not to train on

The decision shows up as a string in a config file. `muse-spark-1.3` and `muse-spark-1.3-contributor` are one word apart, and the word decides whether the text your users typed goes on to train future Meta models ^[3][4].

The discount is uneven across the three meters. Standard input costs 12.5 times contributor input ^[1], output 21.25 times ^[2], and cached input 75 times ^[3]. For a job that sends a million input tokens and gets 200,000 back, standard bills $2.10 and contributor $0.14, so Meta's promise not to train on that traffic costs $1.96 per run ^[4].

The rate limits decide who can collect the discount. Meta's documentation says the contributor tier "lowers the barrier to entry for prototyping, testing integrations, and scaling experiments where training on your data is acceptable" ^[5]. To spend the contributor allowance of 3 million tokens a minute you have to average 30,000 tokens per request; on standard, the equivalent average is 1,333 ^[5]. The 100-request ceiling also holds you to 144,000 calls a day ^[6]. That fits batched document work and rules out chat turns and per-row classification. Long prompts bill at the same per-token rate as short ones, so filling the context window is not penalised on its own ^[11].

Web search grounding and audio time sit outside the tier choice. Grounding costs $2.50 per 1,000 queries on top of the request's token cost, listed under the shared pricing notes ^[10]. At contributor rates, one grounded search costs $0.0025 against $0.0006 of tokens for a 5,000-token prompt returning 500 ^[7]. Audio time is billed as a separate line item from token usage, rounded down to whole seconds ^[9].

For Monday, two facts about the workload settle it. The first is whether the text you send is yours to give away, and customer content covered by a contract that forbids training use is not. The second is whether your call pattern survives 100 requests a minute ^[6]. Own text in long batched prompts is the one case where contributor pays; own text in many small calls takes standard, because the request cap bites before the savings arrive. Both customer-text cases land on standard, and the data-rights budget line is the gap priced at your own input-output mix. The contributor list covers the 1.3 and 1.2 builds only, so a team pinned to `muse-spark-1.1` stays on standard pricing ^[4].