In an initiative to compete with companies like OpenAI and Anthropic, Meta has introduced a reduced-cost option for Muse Spark. This model, focused on agentic artificial intelligence and programming, now offers a plan that significantly lowers the cost of tokens, provided Meta can use the supplied data for training purposes.
This new plan, called 'contributor,' offers savings of up to 95% compared to the standard mode, where data remains protected against training processes. With this discount, the cost of processing responses drops to only 5% of the original value.
Meta indicates that this mode is ideal for integration tests, prototyping, and experiments—that is, scenarios where concern over data usage is low. The plan supports Muse Spark 1.3 Contributor and Muse Spark 1.2 Contributor models, aiming to facilitate simpler activities that do not involve confidential customer information or proprietary code.
The strategy aims to stimulate the collection of more data related to developers' real-world activities to train the company's AI agents, a goal that, according to TechCrunch, has been challenging. Previously, Meta had implemented an internal program to monitor employee computer usage to generate training data, but this was suspended in June.
The challenge of obtaining data is even more evident outside the field of programming. Although there are many records on software development, there is a scarcity of practical examples of how users interact with spreadsheets, administrative systems, and other corporate tools in daily life.
In addition to the training benefit, the discount positions Muse Spark in direct competition with AI providers like OpenAI and Anthropic, who have also announced price reductions for their APIs and are seeking to increase the adoption of their models on a large scale.
The cost of AI tokens has become a sensitive point, as various corporations may have implemented internal AI systems without anticipating the subscription model. Unlike common monthly chatbot subscriptions, APIs charge based on token usage, which are text fragments used in processing commands (prompts).
The price reduction coincides with the launch of Muse Spark 1.3 this Wednesday, September 2nd, following a series of monthly updates started in July, including the Muse Spark 1.1 version, which was updated about a month ago.
A central aspect of the company's new model is the savings provided: Meta states that Muse Spark 1.3 uses 25% fewer tokens compared to its previous version. The family of models, intended for more complex tasks and autonomous actions, has also received improvements in multi-task management and the continuity of extensive workflows, in addition to reducing interruptions.
The same innovations were presented by OpenAI with the launch of GPT-6 Astra, which occurred just one day later. The results obtained by Sam Altman's company in autonomous task execution tests (OSWorld 2.0) surpassed those of Meta, achieving 72.6% accuracy against Meta's 66.9%.