Loading page…
The beta agent can modify and check code across large repositories, but its cheapest tier lets Meta train on the prompts and outputs routed through it—a trade teams must assess before plugging in their development work.
Listen to this story
Meta released Muse Code (beta), a command-line coding agent built on Muse Spark 1.2, and launched a discounted 'contributor' tier that charges $0.10 per million input tokens (a 92% cut from the $1.25 standard) in exchange for Meta’s right to train on prompts and outputs routed through the agent. Contributor access is capped at 100 requests/minute/team (vs 3,000 standard).
Contributor price: $0.10 per million input tokens — 92% discount vs $1.25 standard.
Throughput cap: contributor 100 requests/minute/team; standard 3,000 — limits larger-team workloads.
Muse Code is a CLI harness with persistent parallel subagents and local logs to resume interrupted runs.
Meta’s first coding agent arrives as two propositions bound together: a terminal-based system designed to break down work, modify files and check results across large repositories, and a discounted usage tier that gives Meta the right to train on the prompts and outputs passing through it. Muse Code is in beta and runs on the newly released Muse Spark 1.2 model.
Muse Code was built by Alexandr Wang’s team and announced on X by Mark Zuckerberg. Its intended work places it in the same category as Claude Code and Codex: software agents that operate across a repository rather than merely suggesting a line or function. Meta has also said it intends to compete on pay-as-you-go pricing.
The product is a command-line agentic harness, not just a model endpoint. Muse Spark 1.2 supplies the underlying model, while Muse Code coordinates the broader job: forming a plan, editing code and checking the result in a large codebase. That distinction matters because users are buying both the model’s capabilities and the harness that manages a longer-running development process.
Zuckerberg’s description of the architecture says persistent subagents work in parallel inside isolated worktrees. The harness also records every model call and file edit locally. If a run crashes, those records allow it to resume instead of beginning the job again, preserving the work already completed during a long session.
Standard access costs $1.25 per million input tokens. The contributor tier cuts that to $0.10, a 92% discount, in exchange for Meta’s right to use the prompts and outputs moving through the agent for training. The terms described in the available evidence concern information that passes through Muse Code; they do not establish that every file in a connected repository is automatically training material.
Standard access costs $1.25 per million input tokens.
The contributor tier costs $0.10 per million input tokens and grants Meta training rights over prompts and outputs passing through the agent.
The contributor price represents a 92% discount from standard access.
The cheaper tier also has a much lower throughput ceiling: 100 requests per minute per team, compared with 3,000 under standard access. The published account describes that cap as limiting the contributor option largely to individuals and small teams. The offer is therefore not only a data bargain; it is also a capacity trade-off that may rule out heavier workloads.
Muse Spark 1.2 has an eye-catching reported result on Vals AI’s Finance Agent v2, a benchmark that assigns models work resembling that of entry-level financial analysts. At maximum reasoning, the model ranked first of 45 models at a reported cost of $0.77 per task. It placed fifth on the broader Vals Index at $0.70 per task.
Those measurements establish reported benchmark position and task cost. They do not, by themselves, measure developer throughput inside Muse Code: the benchmark evaluates Muse Spark 1.2, while the released product adds planning, file editing, parallel subagents and recovery behavior through its harness. Teams still need to test the combined system against their repositories, task mix and review process.
Separate research summarized by Forbes shows why productivity claims deserve caution. In a randomized trial, experienced open-source developers took 19% longer on real tasks when allowed AI tools, even though they later estimated that the tools had made them 20% faster. METR ended that design in February 2026 because it could no longer recruit developers willing to join a control group without agents.
A replacement analysis of transcripts from seven staff engineers produced estimated time savings ranging from 1.5 times to 13 times, but its author cautioned that the true productivity multiplier was substantially lower. The two approaches do not supply a universal number for coding-agent gains; they instead show how strongly the answer can depend on study design, task selection and measurement.
Meta’s surrounding releases add another layer to the decision. The company released Muse Glimmer, a compact model with available weights that operates on the user’s own device and is small enough to run on a laptop. Meta has also said it plans to publish Muse Spark 1.2’s weights, although that step had not been completed in the available evidence.
Zuckerberg paired the Glimmer release with an essay arguing for broad access to superintelligence. At the same time, Muse Code’s deepest price cut is tied to training rights over service interactions. These choices operate at different layers—access to model weights on one side, treatment of prompts and outputs on the other—and teams should evaluate them separately rather than treating an open-weight plan as a complete answer to service-level data governance.
Andrew Ng’s OpenWorker illustrates a different configuration. The MIT-licensed desktop agent runs locally, works across everyday applications and accepts roughly 30 models through API keys supplied by the user. Where Muse Code packages Meta’s model, orchestration and pricing together, OpenWorker emphasizes local operation and model choice.
Loading discussion...
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.