Meta has officially entered the competitive agentic coding market with the release of Muse Code, a terminal-based AI coding agent currently in beta, alongside Muse Spark 1.2, a specialized frontier model update. Available now for macOS and Linux via a terminal installation command, the release pits Meta directly against established tools like Anthropic’s Claude Code and OpenAI’s Codex.
The developer tool market has shifted dramatically toward autonomous coding harnesses, and Meta’s entry brings distinct architectural choices. Rather than relying on ephemeral processes, Muse Code introduces persistent async background agents that stay active throughout a session. This design aims to eliminate redundant codebase exploration by maintaining contextual awareness across user prompts.
When handling larger tasks, Muse Code fans out to parallel sub-agents operating within isolated Git worktrees. Meta co-founder and CEO Mark Zuckerberg noted on X that testing demonstrated successful parallel feature building without repository collisions. Furthermore, the runtime relies on a local append-only event log capturing every model call, tool run, and file edit. This mechanism makes sessions replay-exact and restart-safe, allowing long-running tasks to resume precisely where they left off after an interruption.
Underpinning the harness is Muse Spark 1.2, a coding-focused iteration that scales up training compute and environment diversity compared to its 1.1 predecessor. Meta co-trained the model directly with Muse Code using rejection-sampled harness trajectories and a self-improvement loop where the previous model generated coding environments and graded solutions.
Benchmark data published by Meta illustrates a competitive standing against rival frontier architectures. On Terminal-Bench 2.1, Muse Spark 1.2 running inside Muse Code scored 82.9%, placing ahead of OpenAI’s GPT-5.6 Terra (81.8%) and xAI’s Grok 4.5 (81.6%), though trailing Anthropic’s Opus 5 at 86.7%. On DeepSWE 1.1, the model registered 59.3%, sitting behind Opus 5 (65.0%) and GPT-5.6 Terra (64.8%). In Meta’s internal coding evaluations, the model achieved 70.6%, outperforming GPT-5.6 Terra (65.4%) but remaining behind Anthropic’s flagship model.
API Pricing Strategy and the Contributor Tier
The commercial structure accompanying Muse Spark 1.2 introduces a sharp divergence from traditional cloud API models. Meta provides two distinct tiers through the Meta Model API:
- Standard Tier: Priced at $1.25 per million input tokens and $4.25 per million output tokens, with cached inputs at $0.15. Prompts and completions on this tier are explicitly excluded from model training.
- Contributor Tier: Priced at $0.10 per million input tokens and $0.20 per million output tokens, with cached inputs at $0.002. Users on this tier grant Meta explicit permission to use their prompts and completions to train future models.
While the contributor tier offers aggressive cost reductions, it imposes strict rate limits of 60 requests per minute—compared to 3,000 requests per minute on the standard tier—and requires a payment method on file before execution. For engineering teams handling proprietary or legally protected source code, the standard tier remains the necessary choice to prevent codebases from entering Meta’s training pipeline.
The Absence of Open-Source Artifacts
Noticeably absent from the release is any provision of open-weight models or self-hosting licenses. For years, Meta positioned the Llama family as the standard for open-source AI, accumulating roughly 1.2 billion downloads. However, following the restructuring into Meta Superintelligence Labs and the debut of the original Muse Spark in April as a proprietary, cloud-only architecture, the company has pivoted away from releasing downloadable weights.
While Zuckerberg responded to inquiries on social media indicating that open-source releases might return in the future, today’s launch features no open weights or permissive licensing. This places Muse Code and Muse Spark 1.2 closer to Anthropic’s proprietary posture than to open-source alternatives offered by competitors.
Enterprise Integration and Outlook
Meta demonstrated long-horizon capabilities by pointing Muse Spark 1.2 at GPU kernel optimization tasks on NVIDIA Hopper hardware, executing over 1,000 tool calls across 24 hours. Working within Triton, the agent generated non-obvious optimizations for KDA and MLA kernels.
As enterprise engineering leaders evaluate terminal-based workflows, adoption will likely depend on real-world repository performance, trust in data handling policies across the API tiers, and whether the parallel agent architecture consistently prevents regressions during complex software engineering tasks.
