Meta has entered the AI coding agent arena with Muse Code, a terminal-based tool now in beta that targets professional developers working across large repositories. Announced on August 5, 2026, the agent is powered by Muse Spark 1.2, a coding-specialized update to Meta's frontier model family, and marks the company's most direct challenge yet to Anthropic's Claude Code and OpenAI's Codex. The launch comes from Meta Superintelligence Labs, led by AI chief Alexandr Wang, who joined the company in June 2025 as part of CEO Mark Zuckerberg's push to overhaul Meta's AI strategy.
Muse Code distinguishes itself through a combination of persistent async background agents, parallel sub-agent execution in isolated git worktrees, and a local event log that makes runs replay-exact and restart-safe. Developers can install it on macOS or Linux with a single curl command, though a Meta account and billing details are required. Pricing is structured in two tiers: a standard pay-as-you-go option at $1.25 per million input tokens and $4.25 per million output tokens, and a heavily discounted contributor tier at $0.10 and $0.20 respectively in exchange for permission to use prompts and completions for future model training.
What's New / Specs
- Muse Code (beta): Terminal coding agent for macOS and Linux; installable via curl -fsSL https://dev.meta.ai/install.sh | bash.
- Core capabilities: Planning changes, writing code, validating results across large repositories; coordinates multiple persistent sub-agents for complex tasks.
- Async background agents: Specialized agents remain active throughout a session, avoiding redundant repository exploration and reducing latency on multi-step tasks.
- Parallel worktree isolation: Large jobs fan out to separate sub-agents, each in its own isolated git worktree; the developer's working copy is never touched. Zuckerberg noted testing where six game features were built simultaneously with no collisions.
- Local event log: Every model call, tool run, approval, and edit is appended before execution, enabling replay-exact and restart-safe runtime — critical for long-running tasks that may span 24+ hours.
- Bundled skills: /plan creates approval-gated plans, /grill stress-tests plans, /goal drives toward objective completion.
- Muse Spark 1.2: Coding-focused update with significantly scaled training compute on coding tasks, expanded training environment diversity, and co-training with Muse Code using rejection-sampled harness trajectories and recipe optimizations for goals, context compaction, and sub-agents.
- Self-improvement loop: Muse Spark 1.1 generated challenging coding environments and instruction-following templates, then graded candidate solutions to produce a scalable training dataset for 1.2.
- Benchmark results (Meta-reported): Terminal-Bench 2.1: 82.9% (2nd behind Anthropic Opus 5 at 86.7%); DeepSWE 1.1: 59.3% (3rd behind Opus 5 at 65.0% and GPT-5.6 Terra at 64.8%); Internal coding benchmark: 70.6% (beats GPT-5.6 Terra 65.4% and Gemini 3.6 Flash 63.9%, trails Opus 5 at 79.4%).
- Case study: GPU kernel optimization over 1,000+ tool calls, up to 24 hours on NVIDIA Hopper GPUs; achieved substantial improvements on KDA and MLA kernels in Triton without wrapping third-party libraries.
- Pricing tiers: Standard — $1.25/M input, $4.25/M output (cached $0.15/M), no long-context premium, 3,000 req/min and 4M tokens/min per team; Contributor — $0.10/M input, $0.20/M output (cached $0.002/M), requires opt-in for data use in training.
- Availability: Muse Code beta and Muse Spark 1.2 via Meta Model API and OpenRouter; zero-data-retention option for enterprise customers.
Muse Spark 1.2 maintains strength in general agentic tasks while improving code generation, complex debugging, and codebase understanding. The co-training approach means the model was explicitly tuned to perform best inside the Muse Code harness, mirroring an industry shift toward tight model-harness integration. Meta's official blog notes that larger and more capable models are on the horizon, along with new harness features.
The contributor tier pricing — roughly 12x cheaper on input and 21x cheaper on output than standard — is the most aggressive price-positioning in the market. However, it comes with a clear trade-off: developers must explicitly opt in to allow Meta to use their prompts and completions to train future models. Wang characterized this as a way to "help improve the model" and said adoption of Muse Spark models has been "exciting and strong," though he declined to share specific user statistics.
Why It Matters
Meta's entry reshapes the competitive dynamics of AI-assisted software development. Until now, the category has been defined by Anthropic's Claude Code and OpenAI's Codex, with startups like Cursor building billion-dollar businesses on top of those foundations. Meta's developer story previously centered on Llama, the open-weight model family with over a billion downloads, but Muse Code and Muse Spark are entirely proprietary — a notable strategic pivot. Zuckerberg teased on X that open source may return for future Muse products, saying "I'll have more to share on that soon." The combination of a purpose-built harness, co-trained model, and aggressive pricing signals Meta's intent to compete for enterprise and professional developer mindshare directly.
The technical architecture addresses persistent pain points in agentic coding. Persistent background agents reduce the re-exploration penalty that plagues stateless agent designs, while worktree isolation eliminates the collision risk when multiple agents modify the same codebase. The local event log provides an audit trail that enterprise evaluators have demanded: every decision is recorded before execution, enabling exact replay and crash recovery without lost work. These features, combined with the 24-hour kernel optimization demonstration, suggest Meta is targeting sustained, long-horizon autonomy rather than quick snippet generation.
Pricing strategy is the other lever. The standard tier sits mid-range among frontier model APIs, but the contributor tier undercuts every competitor by an order of magnitude. For cost-sensitive teams — startups, open-source maintainers, academic labs — the $0.30 per million total token cost is compelling. The trade-off is data: Meta gains a continuous stream of real-world coding interactions to improve future models. This mirrors the data flywheel that has powered Meta's advertising business, now applied to model training. Enterprise customers wary of data exposure can opt for zero-data-retention, a feature Wang called "a big enterprise feature that is important for folks." Muse Spark 1.2 is also available on OpenRouter, expanding distribution beyond Meta's own API.
Benchmarks tell a nuanced story. Muse Spark 1.2 shows clear generational gains — +6.7 points on Terminal-Bench and +6.3 on DeepSWE over 1.1 — but the charts also reveal a consistent second-place position behind Anthropic's Opus 5 across all three reported evaluations. On Meta's own internal benchmark, Opus 5 leads by nearly nine points. The company's decision to publish these comparisons without claiming the top spot is unusual in a field where leaderboard wins are routinely trumpeted. It suggests confidence in the product-market fit of the harness-model combination rather than raw model supremacy.
Our Take
Meta's launch is credible and technically substantive. The harness architecture — persistent agents, worktree isolation, replay-exact logging — solves real problems that developers encounter daily with current tools. The co-training methodology is sound: a model optimized for its own toolchain will outperform a generalist model dropped into the same harness. The kernel optimization case study, if reproducible outside Meta's controlled demos, demonstrates a level of sustained autonomous improvement that most coding agents fail to achieve past the first few hundred tool calls.
The pricing strategy is the boldest move. The contributor tier effectively subsidizes model improvement with user data, creating a flywheel that could accelerate Muse Spark's trajectory faster than rivals relying solely on curated datasets. But it also introduces a trust dependency: teams must weigh the 12-21x cost savings against the intellectual property implications of sending proprietary code through a training pipeline. Zero-data-retention mitigates this for enterprises, but the contributor tier's terms deserve scrutiny — especially regarding what "opt-in" covers and whether it can be revoked retroactively.
Benchmark transparency is refreshing but incomplete. The charts compare against strong contemporaries (Opus 5, GPT-5.6 Terra, Grok 4.5, Gemini 3.6 Flash), yet the 1.1 baseline ran in a generic mini-swe-agent harness while 1.2 ran in Muse Code, conflating model and harness gains. A clean ablation — 1.2 in the old harness, 1.1 in the new — would clarify how much progress belongs to each. The internal benchmark, while favorable to Meta's design choices, still shows a nine-point gap to Opus 5. For teams evaluating solely on SWE-Bench-style metrics, Anthropic remains the leader.
The open-source question hangs over the launch. Llama's success was built on open weights and community adoption; Muse Spark and Muse Code are proprietary. Zuckerberg's hint at future open-source releases suggests a staged strategy: establish product-market fit and revenue traction first, then open components to drive ecosystem adoption. Until that happens, Meta is asking developers to bet on a closed platform — a harder sell than the open alternative it once championed.
FAQ
What is Muse Code and how do I install it?
Muse Code is a terminal-based AI coding agent in beta that handles complete software engineering tasks across large repositories. It runs on macOS and Linux and can be installed with a single command: curl -fsSL https://dev.meta.ai/install.sh | bash. A Meta account and billing details are required before use.
How does Muse Code differ from Claude Code or OpenAI Codex?
Muse Code uses persistent async background agents that stay active throughout a session rather than spawning fresh for each task, reducing redundant repository exploration. It executes large jobs via parallel sub-agents in isolated git worktrees so the developer's working copy is never touched. A local event log records every model call, tool run, and edit before execution, enabling exact replay and crash recovery. It also ships with built-in skills like /plan, /grill, and /goal.
What are the pricing tiers and what is the catch with the contributor tier?
The standard tier costs $1.25 per million input tokens and $4.25 per million output tokens (cached input at $0.15), with prompts and completions not used for training. The contributor tier costs $0.10 per million input and $0.20 per million output (cached at $0.002) — roughly 12x and 21x cheaper — but requires explicit opt-in to allow Meta to use your prompts and completions to train future models.
How does Muse Spark 1.2 perform on coding benchmarks?
On Terminal-Bench 2.1, Muse Spark 1.2 in Muse Code scores 82.9%, behind Anthropic Opus 5 at 86.7% but ahead of GPT-5.6 Terra (81.8%) and Grok 4.5 (81.6%). On DeepSWE 1.1 it scores 59.3% (third behind Opus 5 at 65.0% and GPT-5.6 Terra at 64.8%). On Meta's internal coding benchmark it scores 70.6%, beating GPT-5.6 Terra (65.4%) and Gemini 3.6 Flash (63.9%) but trailing Opus 5 at 79.4%. Generational gains over Muse Spark 1.1 are +6.7 points on Terminal-Bench and +6.3 on DeepSWE, though some improvement comes from the new harness rather than the model alone.
Is Muse Code open source? Will Muse Spark weights be released?
No, Muse Code and Muse Spark 1.2 are proprietary. This is a departure from Meta's Llama strategy. However, Mark Zuckerberg replied on X that open source may be in the cards for Muse Spark or another product, saying "I'll have more to share on that soon." No timeline or commitment has been provided.
