Meta is offering developers a discount of roughly 95% on its new Muse Spark model, an AI system built for coding and other agentic tasks, in exchange for allowing the company to use their prompts and outputs to train future models. Under the “contributor” pricing tier, input tokens drop from $1.25 to 10 cents per million, while output tokens fall from $4.25 to 20 cents per million compared to standard pricing.
The move comes after Meta faced setbacks in gathering training data, including the pause of an internal employee computer-tracking initiative earlier this year following internal criticism. The company did not respond to a request for comment on the new pricing structure.
Developer Mario Zechner, creator of the open-source harness Pi, has previously said access to real coding session data has been central to major capability gains in AI coding agents. Princeton professor Arvind Narayanan noted that large companies have generally avoided sharing data with model providers, often opting for costlier enterprise plans specifically to retain data privacy, suggesting Meta’s discount could push companies to reassess which data is genuinely sensitive versus shareable.
The pricing approach also reflects broader cost competition among AI labs, following recent price adjustments from both Anthropic and OpenAI on their latest models.
