Meta Muse Code

Overview

Meta Muse Code is Meta’s coding agent, built on the Spark 1.2 model and launched in August 2026. It enters a crowded field — Claude Code, Cursor, Codex, Aider — but differentiates on pricing structure rather than raw capability. The core pitch: let Meta train on your code, and your costs drop by an order of magnitude.

Standard pricing puts output tokens at $11.50/Mtok and input at $4.15/Mtok, roughly competitive with frontier coding models. The training-opt-in tier cuts those to $1.15/Mtok output and $0.40/Mtok input — a 10x and ~10x discount respectively. For teams running heavy agentic workloads where token costs are a real line item, the math is straightforward: the discount funds itself within days of moderate usage.

Key Capabilities

Persistent background agents: Muse Code agents can continue executing tasks after you close your editor or terminal. Start a large refactor, close your laptop, and the agent keeps working. Results are available when you reconnect. This is a meaningful workflow improvement over agents that die with the session.

Local event logs and resumability: Every agent action is logged to a local event store. If an agent is interrupted — crash, network drop, manual stop — it can resume from the last checkpoint rather than restarting from scratch. This matters most for long-running tasks like codebase migrations or large test suite generation.

Tool-use loop: The agent operates in a multi-step tool-use loop, reading files, running commands, editing code, and verifying results. It supports multi-file refactors and can reason across a codebase rather than operating file-by-file.

The Training Data Trade-Off

The pricing model is the most interesting design decision. Meta is transparent about the exchange: your code becomes training data for future Spark models, and you get significantly cheaper compute in return. For open-source projects or teams without IP sensitivity, this is a compelling deal. For companies with proprietary codebases, trade secrets, or regulated data, the standard tier keeps your code private at market-rate pricing.

The opt-in is granular — you choose which repositories participate — so teams can use the discount tier for open-source work while keeping proprietary repos on standard pricing.

Considerations

Spark 1.2 is Meta’s latest model but has less public benchmarking data than Claude or GPT-series models in coding contexts. Early adopters should evaluate it against their specific language and framework stack before committing to large-scale usage. The training data contribution model also raises questions about code ownership and licensing that teams should review with legal counsel, particularly for repositories with mixed licensing or contributor agreements.

The persistent agent and resumability features are genuine differentiators, but they require trusting Meta’s local runtime with extended access to your development environment. Review the permissions model before deployment.

Who It’s For

Cost-sensitive teams running heavy agentic coding workloads who are comfortable with the training data trade-off. Open-source maintainers who want frontier-class coding assistance at a fraction of the typical cost. Enterprise teams that want persistent background agents and session resumability, and are willing to pay standard rates to keep code private.