Chief Executive Mark Zuckerberg announced the release of Meta's first AI coding agent, Muse Code, in a social media post on Wednesday, marking the company's entry into the competitive AI coding tools market. According to Business Standard, Zuckerberg stated that Meta had released a beta-test version of Muse Code, which is capable of software engineering tasks like writing code and validating results. The announcement positions Meta directly against OpenAI and Anthropic in the rapidly evolving AI coding landscape, as the company seeks to catch up to its competitors' coding capabilities. Meta has been working to improve its AI models' coding abilities, with Zuckerberg noting in a recent interview that employees were using Muse Spark 1.1 to build products and features for various apps.
According to reports from TechCrunch, Meta launched Muse Code in beta on Wednesday, marking the company's first artificial intelligence coding agent. However, Anthropic's Claude Opus 5 emerged victorious in all four comparisons published by Meta at launch. Muse Spark 1.2, the model inside Muse Code, scored 82.9% on Terminal-Bench 2.1, while Claude Opus 5 achieved 86.7% on the same test. Muse Code demonstrated competitive performance by beating OpenAI's Codex at 81.8% and Grok Build at 81.6%. The benchmark charts explain why Meta showed unusual reticence in trumpeting leaderboard wins, as they show a strong but clear second place across all testing scenarios. Meta announced that "We're excited to release Muse Code (beta), a terminal coding agent powered by Muse Spark 1.2, our newest model," marking their next step toward the frontier with larger and much more capable models on the way.
As reported by TechCrunch, independent test data suggests the performance gap between the two systems may be even wider than Meta's published figures indicate. On the harsher DeepSWE 1.1 test, which sets 113 coding tasks with internet access switched off during grading, Muse Spark 1.2 dropped to third place at 59.3%. Meta also published a test built from 440 real pull requests by its own engineers, where Muse Spark 1.2 scored 70.6%, roughly 9 points behind Opus 5. This score sits only 2.3 points above Muse Spark 1.1, the model Meta shipped in July. The most striking admission comes from Meta's own internal coding benchmark, where Muse Spark 1.2's 70.6% comfortably beats GPT-5.6 Terra (**65.4%) and Gemini 3.6 Flash (63.9%) yet still sits nearly 9 points behind Opus 5's 79.4% - an unusually candid admission that even on the test Meta designed itself, Anthropic's model wins. Recent developments show OpenAI's GPT-5.6 Sol has improved to 73% on DeepSWE 1.1, while Anthropic's Claude Opus 5 reached 74% after its July 24 release, highlighting the ongoing competitive pressure.
According to The Information, Meta is implementing a novel approach to improve its AI coding agent by leveraging real engineering mistakes as training data. Maher Saba, vice president of Meta's Applied AI Engineering organization, has asked engineers across the company to submit at least one code "diff" each week using MetaCode, the company's internal AI coding agent. The initiative has already generated significant results, with 7,000 weekly active users submitting more than 800 fixes so far. Meta has even added colored badges to employees' internal profiles based on the number of changes they contribute to encourage finding more improvements. This approach allows Meta to see the entire development process - the original task, MetaCode's response, the engineer's correction, and any tests or reviews needed for approval. The company plans to use these corrections to post-train an upcoming model known internally as Watermelon, representing a deliberate process where engineers submit corrections after MetaCode gets something wrong.
According to TechCrunch, Meta's most striking demonstration is its async background agents architecture, where specialized background agents remain active throughout each session rather than being spawned for individual tasks. This approach avoids redundant information gathering and carries out next steps on their own, choosing when to report back to the main agent. The system includes a local event log that appends every model call, tool run, approval, and edit before execution, making the runtime "replay-exact and restart-safe." If Muse Code crashes 20 hours into a long-running task, it resumes precisely where it stopped, with no lost work and no re-prompting. The tool also ships with bundled skills including /plan for task approval-gating, /grill for stress-testing plans, and /goal for driving toward completion of stated objectives. In testing, Zuckerberg reported that Muse Code built six features for a game simultaneously with no collisions, demonstrating parallel sub-agent workflow. Meta co-trained Muse Spark 1.2 with Muse Code so the core LLM and the agent work together in synergy, with the agent improving over time as tool calls accumulate.
According to TechCrunch, Meta measured itself against GPT-5.6 Terra, which OpenAI sells as Sol, but Meta left Sol out of all three coding charts. Sol tops the independent Terminal-Bench 2.1 leaderboard at 89.5%, while Opus 5 follows at 89.1%. Against Sol, Muse Spark 1.2 trails by 6.6 points rather than the reported 3.8 points. Meta's pricing strategy introduces a significant divergence from open-source models, with the standard tier priced at $1.25 per million input tokens and $4.25 per million output tokens, matching Meta's first paid API launch. The contributor tier offers dramatically lower rates at $0.10 per million input tokens and $0.20 per million output tokens - roughly 12x and 21x cheaper than standard respectively - but requires explicit permission to use prompts and completions for training future Meta models. This represents a fundamental shift from Meta's previous Llama strategy of free weights in exchange for mindshare to a model where cheap tokens subsidize training data access. Meta is late to the agentic coding fight, with OpenAI's Codex already running parallel cloud agents and competitors like DeepSeek building their own rivals to Claude Code.
