Meta Muse Code enters a crowded market with low token pricing
Meta Muse Code has entered the AI coding market with a low-cost pitch and a model built specifically for software work. Meta launched the beta tool on August 5, using its new Muse Spark 1.2 model to help developers write, test and debug code.
An AI coding agent goes beyond simple autocomplete. It can inspect a project, plan changes, edit several files, run tests and check whether its work succeeded. Meta Muse Code is designed for longer software projects and can run several sub-agents at the same time. Each sub-agent can handle a smaller part of the task while the main system coordinates the result.
According to Reuters, the tool can write code and verify its output. It also keeps an activity log, allowing it to resume work after a crash rather than restart the entire job. Meta said Muse Spark 1.2 and Meta Muse Code were trained together so the model and agent could operate as a pair.
Pricing is central to Meta’s launch. The standard pay-as-you-go tier costs $1.25 per million input tokens and $4.25 per million output tokens. Tokens are small pieces of text or code processed by the model. Input tokens cover the developer’s instructions and project context, while output tokens cover the code and explanations the system generates. Low headline prices can help with trials, but long projects and large codebases may still consume many tokens.
The model follows Meta’s July release of Muse Spark 1.1. In its official model announcement, Meta described the earlier version as a multimodal reasoning model with a one-million-token context window, search with citations and parallel tool use. Muse Spark 1.2 is meant to improve complex instruction following after the earlier model generated and evaluated difficult coding tasks.
Meta Muse Code enters a crowded field. OpenAI offers Codex, while Anthropic has built strong developer adoption around Claude Code. The AI Decode has covered how AI loops in coding can keep agents working across repeated test-and-fix cycles. It has also examined the cost and benchmark pressure around Claude Opus 5.
The main concern is verification. Coding agents can produce plausible software that contains security flaws, broken assumptions or hidden maintenance costs. Teams handling private repositories will also need clear rules for what data Meta stores and whether any discounted access involves sharing development activity.
Meta Muse Code now has to prove that its low price survives real workloads. Developers will be watching reliability, privacy terms, integration quality and whether Muse Spark 1.2 can handle large projects without creating more review work than it saves.
