Meta has unveiled its first coding agent, Muse Code, along with the optimized model Muse Spark 1.2, aiming to compete with Anthropic and OpenAI. The tool performs well on several benchmarks but still trails Claude Code overall. Meta also differentiates through lower pricing and a 'data tax' discount to attract users.
SynthePulse Insight · AI deep reading
Version 1 · 1 source
Meta launches its first coding agent Muse Code and model Muse Spark 1.2, offering discounts in exchange for a 'data tax' at prices below DeepSeek, but benchmarks still lag Anthropic, and open source commitments are absent, raising strategic concerns.
On August 6, 2026, Meta launched its first coding agent Muse Code, along with the programming-optimized model Muse Spark 1.2. CEO Zuckerberg announced on X that Muse Code can 'execute complete software engineering tasks in large codebases,' directly challenging Anthropic and OpenAI. The tool was developed under the leadership of Meta AI head Alexandr Wang, who joined Meta last June.
Muse Code's core architecture is an 'asynchronous background agent': unlike competitors that spawn temporary helper agents, it maintains a set of dedicated background agents alive throughout the session, avoiding redundant information gathering, reducing latency, and minimizing supervision. When tasks are large, they are distributed to independent sub-agents for parallel processing, each operating in an isolated git worktree without touching the developer's working copy.
Auditability is also a design focus: if a task crashes after 20 hours, it can resume precisely from the point of failure without losing work. Every model call, tool run, approval, and edit is appended to a local event log, enabling 'exact replay and restart safety.'
Meta's published benchmarks show Muse Spark 1.2 scoring 82.9% on Terminal-Bench 2.1, slightly above GPT-5.6 Terra (81.8%) and Grok 4.5 (81.6%), but behind Opus 5's 86.7%. On DeepSWE 1.1, Muse Spark 1.2 scores 59.3%, ranking third behind Opus 5 (65.0%) and GPT-5.6 Terra (64.8%).
On Meta's own internal coding benchmark, Muse Spark 1.2 scores 70.6%, beating GPT-5.6 Terra (65.4%) and Gemini 3.6 Flash (63.9%), but still trailing Opus 5's 79.4%. Claude tops all three charts.
Generational improvement is clear: Muse Spark 1.2 improves by 6.7 percentage points on Terminal-Bench over 1.1, and by 6.3 points on DeepSWE. Meta also showcased a long-cycle case: running for 24 hours on NVIDIA Hopper hardware with over a thousand tool calls for GPU kernel optimization, achieving 'substantial improvements.'
Meta offers two API tiers for Muse Code. The standard tier is pay-as-you-go at $1.25 per million input tokens and $4.25 per million output tokens, lower than Anthropic Sonnet 5 (input $3, output $15), with prompts and generated content not used for training, and rate limits of 3000 requests per minute per team and 4 million tokens.
The contributor tier is 'over 10x cheaper than the pay-as-you-go tier,' at $0.10 per million input tokens and $0.20 per million output tokens, even below DeepSeek-V4-Flash (¥1/M input on cache miss, ¥0.2/M on hit, ¥2/M output). However, developers must 'opt in to help improve the model,' allowing Meta to use third-party data to enhance technology, with strict rate limits (60 requests per minute), aimed at individuals and small experiments.
A zero data retention option is available to enterprises at an additional cost, which Meta calls 'a significant enterprise-grade feature important to corporate customers.' This move comes amid Meta's weak earnings last week and a sharp stock drop, with 98% of revenue from online ads; Muse Code is Zuckerberg's new avenue for monetizing AI.
The most notable absence in this release is the lack of mention of 'open source.' Zuckerberg responded to developers on X, saying 'I'll have more to share soon,' hinting at possible future open source plans. For three years, Meta has been the standard-bearer for open AI with the Llama series, with the Llama family downloaded about 1.2 billion times by early 2026, and self-hosting saving enterprises up to 88% in costs.
The rift began with Llama 4: launched in April 2025 to lukewarm reviews, and admitted to inflated benchmarks, while Chinese open source models accounted for about 41% of Hugging Face downloads by the end of 2025. Zuckerberg reorganized AI operations into Meta Superintelligence Labs (MSL) in summer 2025, bringing in Alexandr Wang.
On April 8 this year, MSL released its first proprietary model, Muse Spark, with no downloadable weights, available only via cloud API. Four months later, Muse Code's release still lacks open source commitments, while OpenAI has open-sourced Codex CLI and gpt-oss, and Google's Gemini CLI uses an Apache license. Meta is shifting toward Anthropic's proprietary stance, with the contributor tier becoming the successor to the Llama strategy: trading cheap tokens for training data.
This article is based on a report from AI Frontline; all data and claims come from that single source and have not been independently verified. Benchmarks are self-published by Meta and may be biased. Price and model comparisons are as reported and subject to official confirmation.
Meta enters the AI coding market with low prices and a 'data tax' strategy, but technology still lags Anthropic, and the absence of open source commitments may undermine its developer community foundation. Whether it goes open source in the future will determine its strategic direction.
Primary report
Primary source