By admin 
Meta has made a significant strategic pivot today with the release of Muse Code, a terminal-based AI coding agent now available in beta, and Muse Spark 1.2, an advanced update to its frontier model family. This dual launch positions Meta directly against established players like Anthropic’s Claude Code and OpenAI’s Codex, signaling a departure from its long-standing commitment to open-source AI and a move towards proprietary, data-driven development. The move signifies Meta’s most aggressive foray yet into a burgeoning market where agentic coding tools have rapidly become indispensable for professional developers.
Meta CEO Mark Zuckerberg announced the beta release of Muse Code via a post on X (formerly Twitter), stating, "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." This announcement underscores a clear ambition to compete at the highest level of AI-assisted software development, a space where Meta has largely observed from the sidelines while competitors solidified their positions.
While rivals like Anthropic and OpenAI have integrated their coding agents into flagship product offerings, and startups such as Cursor have built substantial businesses around this workflow, Meta’s previous developer outreach was primarily centered on its Llama family of open-weight models. These models, distributed freely with over a billion downloads, fostered a strong community and championed the philosophy of open AI. Muse Code, however, represents a stark contrast. It is a fully proprietary, installable harness for macOS and Linux, co-trained with its underlying model and available exclusively through Meta’s own infrastructure. Developers can access Muse Code via a single curl command: curl -fsSL https://dev.meta.ai/install.sh | bash. However, access requires a Meta account and the provision of billing details, indicating a commercial-first approach.
A key architectural innovation of Muse Code lies in its implementation of "async background agents." Unlike many competing harnesses that instantiate new agents for each task, Muse Code maintains a pool of specialized background agents that remain active throughout a user’s session. Meta explains in its blog post that these persistent agents "avoid redundant information gathering" and can autonomously execute subsequent steps, reporting back to the main agent when appropriate. This persistent architecture aims to reduce latency and minimize developer oversight, as agents retain knowledge of the repository, eliminating the need for repeated context re-establishment. For larger tasks, Muse Code intelligently distributes work across parallel sub-agents operating within isolated Git worktrees, ensuring the developer’s primary working copy remains untouched. Zuckerberg highlighted this capability, noting, "In testing we had it build six features for a game simultaneously with no collisions." While parallel processing and worktree isolation are features found in other tools, Meta emphasizes the synergy between persistent agents and parallel execution as a unique differentiator.
Another significant design choice is Muse Code’s focus on auditability. Every action, including model calls, tool executions, approvals, and code edits, is logged to a local event log before execution. Meta asserts this creates a "single source of truth" that ensures the runtime is "replay-exact and restart-safe." This means that if Muse Code encounters an issue during a lengthy task, it can resume precisely from the point of interruption without data loss or requiring re-prompting. For engineering leaders concerned about the predictability and transparency of AI agent operations, this comprehensive local audit trail could be a decisive factor in enterprise adoption. Muse Code also includes pre-built "skills" familiar to users of similar tools, such as /plan to generate an approval-gated task breakdown, /grill to rigorously test a plan, and /goal to drive the agent towards completing a specified objective.
Powering Muse Code is Muse Spark 1.2, a version specifically optimized for coding tasks. Meta reports that this update involved "significantly scaled up training compute on coding tasks" and enhanced diversity in its training environment. The result is improved performance in code generation, complex debugging, and codebase comprehension, while retaining general agentic capabilities. This release directly addresses a perceived weakness in the earlier Muse Spark models. When Muse Spark initially launched in April, it demonstrated strong performance in frontier reasoning and vision benchmarks but lagged behind competitors in agentic coding evaluations. For instance, it scored 77.4 on SWE-Bench Verified, falling short of Claude Opus 4.6’s 80.8 and Gemini 3.1 Pro’s 80.6, and also underperformed GPT-5.4 on long-horizon tasks. The coding-specialized checkpoint and the purpose-built harness of Muse Code appear to be Meta’s strategic response to close this gap.
Two key training aspects of Muse Spark 1.2 are particularly noteworthy. Firstly, Meta co-trained the model directly with Muse Code, leveraging rejection-sampled harness trajectories and optimizing for specific goals, context compaction, and sub-agent coordination. This approach, where the model is explicitly tuned for optimal performance within a particular tool, reflects a broader industry trend of integrating models and their harnesses as inseparable products. Secondly, Meta employed a self-improvement loop. Muse Spark 1.1 was used to generate complex coding environments and instruction-following templates. Candidate solutions were then graded against these requirements, creating a scalable training dataset for Muse Spark 1.2. Meta attributes the model’s enhanced ability to follow intricate instructions to this iterative process.
Meta has published benchmark charts comparing Muse Spark 1.2 against other coding models on Terminal-Bench 2.1, DeepSWE 1.1, and an internal Meta coding benchmark. While the announcement refrains from claiming leaderboard victories, the charts reveal a strong, albeit second-place, performance. On Terminal-Bench 2.1, Muse Spark 1.2 achieved 82.9%, surpassing OpenAI’s GPT-5.6 Terra in Codex (81.8%) and xAI’s Grok 4.5 in Grok Build (81.6%). However, it trailed Anthropic’s Opus 5 in Claude Code, which led with 86.7%. On DeepSWE 1.1, Muse Spark 1.2 scored 59.3%, ranking third behind Opus 5 (65.0%) and GPT-5.6 Terra (64.8%). Meta’s internal benchmark showed Muse Spark 1.2 at 70.6%, outperforming GPT-5.6 Terra (65.4%) and Gemini 3.6 Flash (63.9%), but still significantly behind Opus 5’s 79.4%. This candid admission, even on Meta’s own benchmark, highlights Anthropic’s continued leadership in coding performance. Nonetheless, generational improvements are evident, with Muse Spark 1.2 showing gains of 6.7 and 6.3 points over 1.1 on Terminal-Bench and DeepSWE, respectively. It is important to note that Muse Spark 1.2 ran within Muse Code, while 1.1 used the generic mini-swe-agent harness, meaning some of the performance uplift can be attributed to the new harness.
A particularly compelling demonstration of Muse Spark 1.2’s capabilities is a long-horizon case study involving GPU kernel optimization. The agent was tasked with optimizing KDA and MLA kernels on NVIDIA Hopper hardware for over 1,000 tool calls spanning up to 24 hours. Crucially, the agent was restricted from simply integrating existing third-party libraries, requiring it to write, compile, and profile its own solutions in Triton. Meta reports that the agent achieved "substantial improvements" over baseline implementations, including non-trivial optimizations such as re-centering gated cumulative decay at a chunk midpoint. Zuckerberg noted, "It kept finding substantial improvements well beyond the initial exploration phase." Such sustained autonomous improvement over extended periods could address a common criticism of coding agents: their tendency to plateau after initial progress.
The pricing structure for Muse Spark 1.2, accessible via the Meta Model API, is perhaps the most contentious aspect of the launch. Meta offers two tiers: a standard tier and a "contributor" tier. The standard tier is priced at $1.25 per million input tokens and $4.25 per million output tokens, with cached input at $0.15. Meta guarantees that prompts and completions on this tier will not be used for model training. Rate limits are generous at 3,000 requests and 4 million tokens per minute per team, placing it in the mid-range of API-based AI model pricing.
The contributor tier, however, introduces a sharp divergence from competitor strategies. It is priced at an exceptionally low $0.10 per million input tokens and $0.20 per million output tokens – approximately 12x and 21x cheaper than the standard tier, respectively, with cached input at a near-negligible $0.002. The steep discount comes at the cost of explicit permission for Meta to use user prompts and completions for training future models. This makes it the cheapest option on the market, but users effectively trade their data for the cost savings. Zuckerberg actively promotes this tier, stating, "It’s easy and low-cost to get started. Install Muse Code with one line and you can start on our contributor tier."
In preliminary testing, the one-line installer for Muse Code functioned as advertised, with a modest download size. However, the agent halted execution, citing the need for payment to complete setup. This indicates that even the discounted contributor tier requires a payment method on file before any operations can commence, meaning it is low-cost but not free. Meta frames this tier as a means to lower the barrier for prototyping and experimentation where data training is acceptable. Nevertheless, it means that the default onboarding path for Muse Code directs developers’ code and prompts into Meta’s training pipeline. Enterprises with sensitive proprietary codebases will need to consciously opt out by selecting the standard pricing. The contributor tier also imposes significantly tighter rate limits (60 requests per minute versus 3,000), clearly positioning it for individual developers and small-scale experiments rather than production workloads. This strategy is characteristic of Meta: subsidize access, harvest data at scale, and leverage it to accelerate development and close competitive gaps. Zuckerberg’s statement that Muse Spark 1.2 is "our next step as we push toward frontier, with larger, more capable models on the way" reinforces this long-term vision. However, for developers and enterprises with legal or ethical obligations to protect their code, this data-for-discount trade-off may be unacceptable.
Notably absent from today’s announcement is any mention of open source, a striking omission from a company that has spent years championing the open-source AI movement. The trajectory began with the February 2023 leak of LLaMA’s weights, which inadvertently fueled the open-source AI revolution. This was followed by Llama 2’s commercially permissive license, the specialized Code Llama, and the ambitious Llama 3.1, launched in July 2024 with Zuckerberg’s manifesto, "Open Source AI Is the Path Forward." Meta’s consistent message was that frontier-class AI models should be freely available for download, self-hosting, and fine-tuning. This strategy proved highly successful, with the Llama family amassing approximately 1.2 billion downloads by early 2026 and offering significant cost savings for enterprises compared to proprietary API providers.
However, this narrative began to unravel. Llama 4, released in April 2025, received mixed reviews and was later revealed to have misrepresented benchmark results. Concurrently, Chinese open-weight models from DeepSeek, Alibaba, and Zhipu AI gained significant traction, capturing a substantial share of downloads on Hugging Face by late 2025 and eroding Llama’s leadership position. This rocky rollout prompted a restructuring of Meta’s AI operations into Meta Superintelligence Labs (MSL) in summer 2025, with Alexandr Wang, co-founder of Scale AI, appointed as Chief AI Officer.
The "Llama era" effectively concluded on April 8th of this year with the release of the original Muse Spark, described by Wang as "the most powerful model that Meta has released." This marked Meta’s first proprietary model, accessible only via the cloud and a private API preview, with no downloadable weights or self-hosting options. When questioned about the future of Llama development at the time, a Meta spokesperson stated only that existing Llama models would remain open source, pointedly avoiding comment on future iterations. Wang himself had previously indicated plans to "open-source future versions" of larger models, but today’s release offers no tangible progress toward that promise, lacking any mention of weights, licenses, or even the term "open" in its official communications.
This reversal is particularly striking given the opposing trend among Meta’s rivals. OpenAI has open-sourced its Codex CLI under the permissive Apache 2.0 license and released its gpt-oss open-weight models. Similarly, Google’s Gemini CLI harness is also Apache-licensed. Meta’s current posture, with Muse Code, closely mirrors that of Anthropic, whose Claude Code remains proprietary. The company that once advocated for open source as the primary path forward now requires developers to pay per token for a model they cannot inspect or to subsidize access with their own data. Viewed through this lens, the contributor tier appears to be the successor to Meta’s Llama strategy: the ecosystem flywheel is now driven by low-cost tokens in exchange for training data, rather than free weights for mindshare.
The significance of this announcement lies in the rapid ascent of terminal coding agents as a critical enterprise AI surface. Until today, this category was largely dominated by Anthropic and OpenAI, with Google and various startups vying for market share. Meta’s entry introduces a genuinely novel architecture with persistent background agents and an append-only local event log, a compelling long-horizon demonstration, and an aggressive pricing strategy. The key unanswered questions pertain to real-world performance: whether Muse Spark 1.2 can truly rival Claude and GPT-class models on complex, real-world code repositories, whether developers will trust Meta with their proprietary code, and whether the contributor tier’s discount is sufficient incentive to overcome data privacy concerns. Muse Code is currently available in beta, while Muse Spark 1.2 is accessible via the Meta Model API with expanded global availability.
