Meta (META.US) is directly challenging OpenAI and Anthropic: it has launched its first AI programming agent, Muse Code, which emphasizes low pricing and "crash recovery" capabilities.

In the increasingly crowded AI programming tools landscape, Meta (META.US) has finally dropped a bombshell. On Wednesday, Meta CEO Mark Zuckerberg officially announced on social media the launch of the company's first AI programming agentMuse Code (beta version), directly challenging OpenAI's Codex and Anthropic's Claude Code. This move signifies a key step for Meta in the generative AI application layer, and represents a new monetization path presented to the public against the backdrop of pressure on its advertising business and exorbitant spending on AI infrastructure. Not the strongest performance, but "reliability" and "affordability" become the breakthrough points Muse Code is not an isolated product; it is powered by the latest foundational model Muse Spark 1.2 from Meta's Super Intelligence Lab. According to multiple benchmark tests released by Meta, Muse Code's programming capabilities still lag behind those of top competitors, but it displays clear differentiated characteristics. In the Terminal-Bench 2.1 benchmark test, the combination of Muse Code and Muse Spark 1.2 scored 82.9%. Although this is less than the 86.7% achieved by Anthropic's Claude Code on Opus 5, it successfully surpasses OpenAIs Codex based on GPT-5.6 Terra (81.8%) and Grok Build (81.6%). On the DeepSWE 1.1 metric that measures the coding ability of agents and on Meta's internal coding benchmark, its scores are still in the chasing position, trailing Opus 5 and Codex by about 5 to 9 percentage points. Since pure performance has not yet topped the charts, Meta has focused its selling points on engineering reliability and extreme cost-effectiveness. Metas Chief AI Officer Alexandr Wang stated in an interview that compared to products from Anthropic and OpenAI, Meta's differentiation strategy is "to achieve differentiation in price rather than capability." The "contributor tier" pricing of Muse Code is highly aggressive, with a price that is less than one-tenth of conventional pay-as-you-go models, but the catch is that developers opting for the lowest price tier must proactively agree to help improve the model, which involves Meta using third-party data to enhance its underlying technology. For highly sensitive enterprise users, Meta has also begun to accept "zero data retention" requests, promising not to retain developers' data for training purposes. Designed for long-duration tasks: Can recover from crashes and run continuously for 24 hours As a programming tool with autonomous decision-making capabilities, Muse Code is specifically designed to handle complex software engineering tasks involving large codebases. It can autonomously complete the full process of planning modifications, writing code, and validating results, while also coordinating multiple persistent sub-agents to process tasks in parallel to enhance speed and accuracy. One of the features most highly regarded by developers is its runtime mechanism. This system records every model call, tool execution, approval, and edit in a local event log, creating a unique and trustworthy data source. This means that if the agent encounters a crash or interruption during prolonged operation, it can recover precisely from the point of interruption without starting over. While competitors have yet to position this as a core selling point, Meta believes this feature is the most attractive for programmers dealing with long-cycle, high-complexity tasks. In a stress test, Muse Code conducted over 1000 iterations of tool calls for kernel optimization on the NVIDIA Corporation (NVDA.US) Hopper GPU, with the entire task lasting 24 hours. Additionally, Muse Code demonstrated powerful multimodal capabilities: in one demonstration, the user simply entered a drone video file of a house, and Muse Code was able to interpret the video content and directly generate a visually rich website with booking functionality. AI strategic restructuring: Led by Alexandr Wang, eager to prove monetization capabilities The debut of Muse Code is a key outcome of Zuckerberg's restructuring of Meta's AI strategy. Alexandr Wang, who joined last June and heads Meta's Super Intelligence Lab, is at the center of this transformation. Previously, Metas self-developed models were clearly lagging behind ChatGPT and Claude in key capabilities such as programming, a point that Zuckerberg publicly acknowledged last month, revealing that the company was already using a previous version of Muse Spark to build application functionalities. The launch of a programming agent comes at a financially sensitive time for Meta. Last week, Meta's stock price fell due to the company's weak revenue outlook and a significant decline in free cash flow for the second quarter. The market is generally concerned about whether the enormous expenditures it incurs annually on data centers and related computing facilities will yield corresponding returns. The paid API interface provided by Muse Code is a crucial attempt by Zuckerberg to generate direct revenue through AI, beyond the companys advertising business, which accounts for 98% of its income. Wang revealed that developers can currently pay for usage of Muse Code on demand via the API on Meta's developer page, and the model will subsequently also be available on the OpenRouter platform, which hosts popular open-source models like DeepSeek.