Meta has released Muse Code, a terminal-based AI coding agent now in beta, together with Muse Spark 1.2, a coding-focused update to its Muse Spark frontier models. The launches put Meta in direct competition with Anthropic's Claude Code, OpenAI's Codex, and a growing set of agentic coding tools that have become a primary workflow for many professional developers.
Meta CEO Mark Zuckerberg announced Muse Code on X (@finkd): “Releasing Muse Code in beta today. It's a terminal coding agent that takes on complete software engineering tasks across large repos: planning changes, writing code, validating the results.” The release marks Meta's most substantial move into a category it has largely observed from the sidelines.
What Muse Code offers
Muse Code is delivered as a full harness that can be installed on macOS or Linux with a single curl command; Meta notes users must sign in with a Meta account and supply billing information before using it. The installer command shown by Meta is:
curl -fsSL https://dev.meta.ai/install.sh | bash
Two architectural choices are highlighted. First, Muse Code uses "async background agents": instead of spawning helper agents for each task, it keeps a set of specialized background agents alive for the entire session. According to Meta, these agents carry out next steps autonomously and only report back when appropriate, which reduces redundant repository exploration and latency.
For large jobs Muse Code fans out into separate sub-agents working in parallel, each in its own isolated git worktree so a developer's working copy is never modified. Zuckerberg wrote that in testing Muse Code built six features for a game simultaneously with no collisions.
The second design emphasis is auditability: every model call, tool run, approval and edit is appended to a local event log before execution. Meta says this makes the runtime "replay-exact and restart-safe" — if the agent crashes after many hours, it can resume precisely where it left off without lost work or re-prompting. For engineering leaders burned by opaque agent runs, a complete local audit trail could be decisive in enterprise evaluations.
Muse Code ships with bundled "skills" familiar from rival tools: /plan (turns a task into an approval-gated plan), /grill (stress-tests that plan), and /goal (drives the agent to complete an objective).
Muse Spark 1.2: co-trained with its harness
Under the hood Muse Code runs on Muse Spark 1.2, which Meta describes as a coding-focused update to Muse Spark 1.1 with "significantly scaled up training compute on coding tasks" and broader training environment diversity. The update aims to improve code generation, complex debugging, and codebase understanding while retaining general agentic capability.
Two training details stand out. First, Meta co-trained Muse Spark 1.2 with the Muse Code harness, using rejection-sampled harness trajectories and recipe optimizations for goals, context compaction, and sub-agents — explicitly tuning the model to perform best inside this tool. Second, Meta employed a self-improvement loop: Muse Spark 1.1 generated challenging coding environments and instruction-following templates, graded candidate solutions against requirements, and produced a scalable training dataset for 1.2. Meta credits this loop with making 1.2 measurably better at following complex instructions.
Meta published benchmark charts comparing Muse Spark 1.2 to other coding models on Terminal-Bench 2.1, DeepSWE 1.1, and an internal Meta coding benchmark. The charts show a strong but clear second place behind Anthropic's Opus/Claude:
- Terminal-Bench 2.1: Muse Spark 1.2 in Muse Code scored 82.9%, OpenAI GPT-5.6 Terra in Codex 81.8%, xAI Grok 4.5 81.6%, and Anthropic Opus 5 (Claude Code) lead at 86.7%.
- DeepSWE 1.1: Muse Spark 1.2 scored 59.3%, behind Opus 5 at 65.0% and GPT-5.6 Terra at 64.8%.
- Meta internal benchmark: Muse Spark 1.2 posted 70.6%, ahead of GPT-5.6 Terra (65.4%) and Gemini 3.6 Flash (63.9%) but still well behind Opus 5's 79.4%.
Meta reports generational gains: Muse Spark 1.2 improves over 1.1 by 6.7 points on Terminal-Bench and 6.3 on DeepSWE. A caveat is that 1.1's scores were recorded using a generic mini-swe-agent harness while 1.2 ran in Muse Code, so some of the improvement is attributable to the new harness as well as the new model.
Meta highlighted a long-horizon case study: Muse Spark 1.2 optimized GPU kernels with more than 1,000 tool calls and up to 24 hours of autonomous runs on NVIDIA Hopper hardware. Working in Triton and barred from simply wrapping existing third-party kernel libraries, the agent wrote, compiled and profiled its way to what Meta called "substantial improvements" over baseline KDA and MLA implementations, discovering non-obvious optimizations such as re-centering gated cumulative decay at a chunk midpoint. Zuckerberg noted the agent continued finding improvements well beyond the initial exploration phase.
Pricing and the data tradeoff
Muse Spark 1.2 is offered via the Meta Model API in two tiers:
-
Standard tier: $1.25 per million input tokens and $4.25 per million output tokens (with cached input at $0.15). Meta commits that prompts and completions on this tier are not used to train its models. There is no long-context premium, and rate limits are 3,000 requests and 4 million tokens per minute per team.
-
Contributor tier: $0.10 per million input tokens and $0.20 per million output tokens (cached input $0.002). In exchange for much lower prices (roughly 12x and 21x cheaper than standard), users explicitly permit Meta to use their prompts and completions to train future models. The contributor tier also has tighter rate limits (60 requests per minute), signaling it is aimed at individuals and small experiments rather than production workloads.
Zuckerberg is steering new users toward the contributor tier, calling it easy and low-cost to get started. VentureBeat's testing on a Mac mini showed the one-line installer worked (97 MB download and sign-in) but required a payment method on file before any models were visible — even the discounted contributor tier requires billing information.
Meta frames the contributor tier as lowering the barrier for prototyping where training on your data is acceptable, but it also means the default on-ramp sends developers' code and prompts into Meta's training pipeline unless they explicitly choose the standard tier. The pricing approach echoes Meta's broader strategy of subsidizing access while harvesting data to accelerate model improvements.
No Llama weights, no open-source promise
Conspicuously absent from today's announcement is any open-source offering. That marks a sharp break from Meta's earlier positioning as an open-AI standard-bearer. From LLaMA's 2023 debut through Llama 2, Code Llama and Llama 3.1, Meta had encouraged download, self-hosting and fine-tuning; by early 2026 the Llama family had been downloaded roughly 1.2 billion times.
The shift began after Llama 4's troubled launch in April 2025 and other market dynamics; Meta reorganized its AI efforts into Meta Superintelligence Labs (MSL) and recruited Alexandr Wang as chief AI officer. On April 8 (the date MSL shipped the original Muse Spark), Meta released its first proprietary cloud-only model with no downloadable weights and no self-hosting.
At the time Meta said existing Llama models would remain available, but the company has not released weights or even used the word "open" in describing Muse Spark 1.2. That contrasts with peers: OpenAI open-sourced Codex CLI under Apache 2.0 and published gpt-oss open-weight models, while Google's Gemini CLI harness is Apache-licensed. With Muse Code, Meta is closer to Anthropic's proprietary posture and appears to have moved away from its earlier open-source rhetoric.
Why this matters
Terminal coding agents are one of the fastest-growing enterprise AI surfaces, and Meta's entry brings a distinctive architecture (persistent background agents and an append-only local event log), an aggressive pricing wedge, and a long-horizon demo. Open questions remain: whether the model will match Claude- and GPT-class performance on real-world repositories, whether developers will trust Meta with proprietary code, and whether the contributor-tier discount justifies surrendering training data. Muse Code is available in beta today, and Muse Spark 1.2 is live via the Meta Model API with expanded global access.
