Meta has launched a new low-cost AI programming model, Muse Spark 1.2, and its first coding agent, Muse Code. CEO Mark Zuckerberg stated that Muse Code can execute complete software engineering tasks within large codebases, directly challenging AI coding tools from Anthropic and OpenAI. Meta attracts developers with low-cost APIs, but the lower price comes with a condition: users must opt-in to share data to help improve the model.

Just recently, Meta introduced its first coding agent, Muse Code, alongside the optimized programming model Muse Spark 1.2. This marks the company's most serious effort yet to enter the AI coding tool market. Meta CEO Mark Zuckerberg personally announced the news on social media platform X, stating that the beta version of Muse Code can "perform complete software engineering tasks in large codebases," positioning it as a competitor to leading AI labs Anthropic and OpenAI.

Muse Code was developed by Alexandr Wang, Meta's head of AI, who leads the Meta Super Intelligence Lab and oversees foundational model research. Wang joined Meta in June last year as part of Zuckerberg's core personnel restructuring to revitalize the company's struggling AI strategy. The terminal-based programming tool, powered by Muse Spark 1.2, boasts improvements in "code generation, complex debugging, codebase understanding, and end-to-end developer workflows."

Despite its focus on programming, it still lags behind Claude Code. "In testing, we had it build six features for a game simultaneously without any conflicts," Zuckerberg said in a post, emphasizing its parallel processing capability that does not interfere with the developer's working copy. A key architectural highlight of Muse Code is what Meta calls "asynchronous background agents." Unlike most competing tools that generate temporary agents for each task, Muse Code maintains a set of specialized background agents that persist throughout the entire session. These agents "remain active throughout the session rather than being generated temporarily for a single task, avoiding redundant information collection." They autonomously execute subsequent steps and decide when to report back to the main agent.

This architecture results in lower latency and less oversight: an agent that already understands the codebase doesn't need to re-explore it every time the developer makes a new request. For large-scale tasks, Muse Code distributes work to independent sub-agents for parallel processing, each operating in its own isolated Git work tree, so the developer's working copy is never touched. Another notable design choice is auditability. If Muse Code crashes after 20 hours of a long-running task, it can recover precisely from where it stopped without losing work or requiring re-prompting. Every model call, tool run, approval, and edit is appended to a local event log before execution, which Meta calls a "single source of truth" that makes the runtime "precisely replayable and restart-safe."

Notably, the latest Muse Spark model was developed and trained alongside Muse Code, enhancing overall programming performance. Meta released benchmark charts comparing Muse Spark 1.2 with other programming models. On Terminal-Bench 2.1, Muse Spark 1.2 running in Muse Code scored 82.9%, slightly higher than OpenAI's GPT-5.6 Terra (81.8% in Codex) and xAI's Grok 4.5 (81.6% in Grok Build), but trailing behind Anthropic's Opus 5 in Claude Code, which leads at 86.7%. On DeepSWE 1.1, Muse Spark 1.2 achieved 59.3%, ranking third behind Opus 5 (65.0%) and GPT-5.6 Terra (64.8%). Most notably, on Meta's own internal programming benchmark, Muse Spark 1.2's 70.6% easily beat GPT-5.6 Terra (65.4%) and Gemini 3.6 Flash (63.9%), but still fell nearly 9 points behind Opus 5's 79.4%. Even on Meta's own tests, Anthropic's model came out on top. In fact, Claude ranked first across all three charts. However, generational improvements are real: Muse Spark 1.2 improved by 6.7 percentage points over 1.1 on Terminal-Bench and 6.3 points on DeepSWE.

Meta also showcased an impressive long-duration case: allowing Muse Spark 1.2 to run on NVIDIA Hopper hardware for up to 24 hours with over a thousand tool calls for GPU kernel optimization. The agent worked in Triton without being able to encapsulate existing third-party libraries, ultimately achieving "substantial improvements" on KDA and MLA kernels, including non-intuitive optimizations like re-centering gated accumulation decay. Zuckerberg noted, "It continued to discover substantial improvements even after the initial exploration phase."

Pricing strategy: Data for discounts

Affordable pricing appears to be Meta's key way to differentiate Muse Code from popular products from Anthropic and OpenAI. To this end, Meta offers two tiers for its model API. The first is the standard tier, which uses the same pay-as-you-go pricing as Muse Spark: $1.25 per million input tokens and $4.25 per million output tokens. This is lower than Anthropic's Sonnet 5 model, which typically charges $3 per million input tokens and $15 per million output tokens. Meta promises that prompts and generated content in this tier will not be used for training, with rate limits of 3,000 requests per minute and 4 million tokens per team. Wang stated that the agent also has a "contributor tier that offers access at a significantly lower cost," adding it is "even more than 10 times cheaper than the pay-as-you-go tier." This tier costs $0.10 per million input tokens and $0.20 per million output tokens. Notably, this price is even cheaper than DeepSeek-V4-Flash, which charges 1 yuan per million tokens for uncached inputs, 0.2 yuan per million tokens for cached inputs, and 2 yuan per million tokens for outputs. Furthermore, DeepSeek just announced plans to "increase API service pricing overall in the near future, with a significant expected rise." In exchange, developers must "opt-in to help improve the model," meaning Meta uses third-party data to enhance its underlying technology. This tier also has stricter rate limits (60 requests per minute), clearly targeting individuals and small experiments. Meta emphasized in its blog post that a zero-data retention option is available to enterprise customers, meaning the company will not retain developer data to improve the model, but at an additional cost. Wang said this represents "a significant enterprise feature important to business customers."

Last week, the company released disappointing earnings with weak revenue forecasts and shrinking free cash flow, causing its stock price to plummet. Meta generates 98% of its revenue from online advertising, a market it dominates by targeting consumers based on user data. This new coding agent is another avenue for Zuckerberg to generate revenue from AI, while his company continues to invest heavily in data centers and related computing infrastructure.

Shift from open-source strategy?

For Meta, the most notable absence in this release is any mention of "open source." When responding to a developer's question on X, Zuckerberg said, "I'll have more to share soon," seemingly hinting at potential future open-source plans. For three years, Meta positioned itself as a champion of open AI through its Llama series: from the LLaMA leak in February 2023 that sparked a consumer AI boom, to the commercial license of Llama 2, to the 405-billion-parameter Llama 3.1 and Zuckerberg's declaration that "open-source AI is the path forward," Meta consistently promised developers that cutting-edge weights would be free to download, self-host, and fine-tune. This strategy was highly effective—by early 2026, the Llama family had been downloaded approximately 1.2 billion times, averaging a million daily, and self-hosting saved enterprises up to 88% in costs. The cracks began with Llama 4. Launched in April 2025, Llama 4 received mediocre reviews and eventually admitted to inflated benchmark results, while Chinese open-source models like those from DeepSeek, Alibaba, and Zhipu AI rapidly rose, accounting for about 41% of Hugging Face downloads by the end of 2025, eroding Llama's leadership. This setback prompted Zuckerberg to restructure the AI business into the Meta Super Intelligence Lab (MSL) in the summer of 2025, hiring Scale AI co-founder Alexandr Wang. On April 8 of this year, the Llama era effectively ended. MSL released its first proprietary model, Muse Spark, with no downloadable weights, no self-hosting, and only a cloud API. At the time, a Meta spokesperson said only that "current Llama models will continue to be open-sourced," remaining silent on future models. Wang said they "planned to open-source future versions." But four months later, the release of Muse Code comes with no open-source commitment—no weights, no license, and the word "open source" never appears in the press release or Zuckerberg's post. This shift is particularly sharp as competitors move in the opposite direction. OpenAI open-sourced Codex CLI under the Apache 2.0 license and launched the gpt-oss open-weight model; Google's Gemini CLI also uses the Apache license. Meta is now closer to Anthropic's proprietary stance, and the company that once championed open source as the way forward now requires developers to pay per token for model use. The newly introduced contributor tier, in a way, becomes the successor to the Llama strategy: no longer trading free weights for market share, but trading cheap tokens for training data, with users effectively "subsidizing" the model with their own data.