Meta has introduced a lower-priced usage plan for the newly launched Muse Spark model. If users agree to share prompts and model outputs for subsequent model training, the call cost can be approximately 95% lower than the standard plan on average. This also changes the option of "whether to retain user data" from a privacy concern to a clear price exchange.
Sharing data can get you a lower price.
According to the prices published by Meta, under the standard plan, a fee of $1.25 is charged for every 1 million inputs of token; under the "Contributor Pricing" scheme, the price drops to $0.10. The price difference for outputting token is even greater, with the standard plan costing $4.25 for every 1 million and the contributor scheme costing $0.20.
This means that customers who are willing to allow Meta to use their interactive data can test coding agents and other agent applications at significantly lower costs. Meta states in its pricing documentation that this approach helps to lower the barriers to prototype development, integration testing, and experimental scaling, provided that customers agree to have their data used for training purposes.
Real usage records are more valuable.
For large model companies, real usage records are becoming increasingly important, especially in coding proxies and a broader range of agent tools. TechCrunch quotes Pi developer Mario Zechner as saying that a significant improvement in coding proxy capabilities by 2025 is related to Claude Code's default session saving and its use for reinforcement learning training.
As model manufacturers shift their focus to professional workflows beyond software engineering, obtaining such data has become more difficult. Many corporate processes are more complex, and there are fewer traceable digital remnants, which limits model evaluation and iteration.
Enterprises place greater emphasis on data retention.
Princeton University computer science professor Arvind Narayanan stated that current indications suggest that large companies are generally reluctant to allow their data to be used for model training. He mentioned that although subscription plans for consumers are often 10 to 20 times cheaper, many enterprises still choose the enterprise version that is charged at token, with the main differences lying in data retention and the enterprise's IT governance requirements.
Meta previously faced resistance in obtaining training data as well. Reports indicate that the company launched a plan to track employees' computer usage earlier this year, but it sparked internal criticism and was suspended in June. Regarding the new pricing arrangements for Muse Spark, Meta did not respond to TechCrunch's request for comment.
Frontier models continue to engage in price wars
This pricing also reflects that the competition among leading model manufacturers is shifting towards a more detailed cost structure. Just the day before, Anthropic released new models Fable and Mythos, and reduced the processing costs of token. OpenAI also made a significant price cut on its latest model at the end of July.
Against the backdrop of model capabilities converging, prices, data acquisition methods, and the governance requirements of corporate clients are all becoming new variables in the competition on the AI platform.
