Meta's Muse Spark coding model lists at $4.25 per million output tokens — and $0.20 for developers who share their prompts and responses

Meta AI chief Alexandr Wang says the company is setting its new coding tool and the wider Muse Spark family apart from what Anthropic and OpenAI sell on cost instead of on what they can do.

Meta released a beta of Muse Code, an AI coding agent that runs inside the terminal, on August 5. Wang told CNBC that the cheaper tier works out upward of 10 times below even the plan billed purely by usage.

Through the regular API, Muse Spark 1.2 runs $1.25 per million input tokens, $0.15 per million cached input tokens and $4.25 per million output tokens. Developers who consent to Meta using their prompts and responses to improve its products can shift to a tier the company calls Contributor, where cached input falls to $0.01 and output to $0.20. Muse Code itself is software that has to be paid for, runs on macOS and Linux and is built to work tightly alongside Muse Spark 1.2, unlike Meta's open-weight Llama models.

Of the AI models in wide use, DeepSeek's V4-Flash costs the least to run in benchmark testing, which has made the V4 line a price reference point for rival frontier models. Measured against it, Muse Spark 1.2 costs more at standard pricing than both V4-Flash and V4-Pro; only on the discounted Contributor tier does it come in under both for input and output tokens, which places it among the cheapest options for developers willing to share their data.

On capability the case is thinner. Meta's own internal coding benchmark puts Muse at 70.6% against Claude Opus 5's 79.4%, and the agent arrives late in a market that already holds OpenAI's Codex, Anthropic's Claude Code, DeepSeek's rival coding harness and other agentic tools.

The condition attached to the cheap tier narrows who can take it. Meta earns 98% of its revenue from online advertising and leads that market by aiming ads at consumers on the basis of data about users. Wang told CNBC that the company had separately begun taking requests for zero-data retention, under which it would not hold on to developer data to improve its models, and called that a substantial enterprise capability that matters to customers.