Meta launches Muse Code: AI coding agent with aggressive pricing
On Aug. 5, Meta entered the AI coding market with Muse Code, a terminal-based coding agent powered by its new Muse Spark 1.2 model. The launch is notable for two reasons: the pricing is unusually aggressive, and Meta dropped its trademark open source framing entirely.
What Muse Code Is
Muse Code is a standalone terminal agent currently in beta for macOS and Linux. It handles autonomous development workflows, including planning, coding, testing, and result verification in a single pass after receiving a user request.
Two features stand out from the standard AI coding agent playbook. First, persistent asynchronous background agents: rather than spinning up a new helper agent per task, Muse Code keeps specialized sub-agents warm throughout an entire session. According to Meta, this avoids repeated file re-reads that slow down most competing implementations. Sub-agents run inside isolated Git worktrees, so parallel workloads do not contaminate each other’s file state. Branches merge sequentially at integration time, keeping conflict management contained.
Second, append-only local event logs record every model call, tool use, approval, and code edit. If the session crashes, developers can replay the full history and resume from any point. Foreign media outlets cited this as a meaningful step toward proper audit trails and root-cause analysis in agentic coding environments.
Benchmarks and the Controversy Around Them
Meta published benchmark results comparing Muse Spark 1.2 and Muse Code against Grok 4.5, Claude Opus 5, GPT-5.6 Terra, and Gemini 3.6 Flash. On Terminal Bench 2.1, the Muse stack scored 82.9%, ahead of GPT-5.6 Terra and Codex (81.8%) and Grok 4.5 and Grok Build (81.6%), but behind Claude Opus 5 and Claude Code at 86.7%.
The comparison set drew criticism. Developers argued Meta avoided top-tier competitor models in favor of second-tier options, and that a comparison against GPT-5.6 Sol would show lower relative performance. AI tech news site Kingy AI went further, publishing a side-by-side showing Meta’s reported scores as slightly higher than independently verified figures. On Deep SWE 1.1, Muse Spark 1.2 ranked behind both Claude Opus and GPT-5.6 Terra.
️ Pricing: Two Tiers, One of Them Very Cheap
The pricing structure is the most disruptive part of the announcement, according to critics who reviewed it.
- Standard Tier: $1.25 per million input tokens, $4.25 per million output tokens, $0.15 per million cached input tokens. Meta commits not to use prompts or outputs for model training.
- Contributor Tier: $0.10 per million input tokens, $0.20 per million output tokens. In exchange, developers must allow Meta to use their code for future model training.
Meta also confirmed that Zero Data Retention (ZDR) is available to enterprise customers upon request. ZDR means user inputs and outputs are deleted immediately after API call processing, with no temporary server storage at any point. This is typically a hard requirement for regulated industries and organizations with sensitive source code, and its availability on day one signals that Meta is actively targeting enterprise accounts, not just individual developers.
The Context
Meta shipped Muse Code less than a month after releasing Muse Spark 1.1. Observers linked the timing to Moonshot AI’s Kimi K3, which generated significant attention immediately after the Muse Spark 1.1 launch. The absence of any open source framing in the Muse Code announcement is a meaningful shift: Meta’s previous strategy centered on indirect monetization through its open-weight Llama ecosystem. Muse Code is a direct revenue play.
Meta’s Q2 financials add context: revenue of $60.8 billion (up 28% year over year), but operating income down 8% to $18.8 billion, operating margin down from 43% to 31%, and CAPEX of $31.08 billion for the quarter alone, with full-year guidance up to $145 billion. The pricing aggression makes more sense when you see those cost numbers.
The current limitation worth noting: Muse Code is terminal-only at launch. Claude Code supports terminal, IDE, desktop, and web interfaces. OpenAI’s Codex spans CLI tools, IDE extensions, and cloud services. If you need anything outside a terminal today, Muse Code is not yet the answer.
I’m founder of BizStack at Cagri Sarigoz LLC and a passionate advocate for entrepreneurs.
With over 14 years in tech, marketing, and AI, including my role as Head of SEO at CitizenShipper and co-founder of TaleBot at Intale AI, I’m dedicated to sharing genuine, useful product insights and tips.
At BizStack, I aim to cut through the digital noise to provide clear, actionable advice.
And more than all else, I’m a father to a (always) little girl and a husband.
Contact me at [email protected] for assistance.
