Muse Code

Muse Code is a terminal coding agent developed by Meta Superintelligence Labs. Meta introduced it in beta on August 5, 2026 alongside the Muse Spark 1.2 model that powers it ^[1][3], then said it had left beta on August 31 ^[16][17]. It is Meta's first coding agent, entering a category defined by Anthropic's Claude Code, OpenAI's Codex, and Google's Gemini CLI ^[5][9]. Meta pitches the tool at long-horizon software engineering across large repositories: it plans changes, writes code, validates the results, and coordinates persistent sub-agents that work on multi-file tasks with less human steering ^[1][3].

The post-beta release added communication between local sessions, reusable multi-agent Workflows, a rewind control, and a TypeScript SDK in developer preview ^[16][18][19]. Meta also began rolling out monthly subscriptions while retaining usage-based access through the Meta Model API ^[16][19][20]. The company has emphasized price, including a discounted Contributor API tier for developers who permit Meta to use their prompts and completions to improve its products ^[5][9][20]. Muse Spark 1.3 arrived in Muse Code on September 2, 2026, and its max reasoning mode followed on September 4 ^[22][23][24].

Background

Meta Superintelligence Labs, the unit Mark Zuckerberg formed in June 2025 around former Scale AI chief executive Alexandr Wang, replaced the open-weight Llama line with the proprietary Muse Spark family on April 8, 2026 ^[7][15]. Muse Spark 1.1 followed on July 9, 2026, with stronger agent and coding abilities, a 1 million token context window, and the public preview of the Meta Model API for developers in the United States ^[11][12]. TechCrunch framed that release as Meta's entry into "the crowded AI coding battle," and described the company as a straggler in agent harnesses that was trying to catch up ^[6][12].

Muse Code is the harness half of that effort. Meta says Muse Spark 1.2 and Muse Code were co-trained so the model performs best inside its own agent, using rejection-sampled harness trajectories and training recipes tuned for goals, context compaction, and sub-agent delegation ^[1]. CNBC reported the launch as Meta's first coding agent and the latest major release under Wang, who oversees foundation model development ^[5].

Launch

Meta announced Muse Code on August 5, 2026 in a Meta Superintelligence Labs blog post titled "Introducing Muse Code and Muse Spark 1.2" ^[1]. Zuckerberg wrote on X that the tool "takes on complete software engineering tasks across large repos: planning changes, writing code, validating the results" ^[4]. The AI at Meta account described "persistent sub-agents" solving complex multi-file changes ^[3]. The beta installs on macOS and Linux with a single shell command; there is no desktop application, unlike the app-based options that exist for Claude Code and Codex ^[1][8].

The launch drew immediate attention: the announcement reached the front page of Hacker News, where the thread collected 326 points and 257 comments in its first days ^[14].

Post-beta release

On August 31, 2026, Meta said Muse Code had left beta. The release announcement listed inter-session messaging, Workflows, rewind in the command-line interface, an SDK in developer preview, and monthly subscription plans ^[16][17]. At that point Meta's product page identified Muse Spark 1.2 as the model powering Muse Code ^[20]; an archived copy of the page from September 4 still listed "Access to Muse Spark 1.2" in its subscription-plan descriptions even though its pricing table had by then added muse-spark-1.3 rows ^[25]. Independent coverage by The New Stack reported the release and described the subscriptions as additions alongside the existing pay-as-you-go options ^[19].

Inter-session messaging lets sessions owned by the same user on one machine communicate through a Unix socket. Meta says these messages remain local rather than crossing the network ^[16]. Workflows coordinate parallel agents, expose their progress through the /workflows control view, and can save completed workflows for reuse; the published release changelog described Workflows as Linux-only at launch ^[16][21]. Rewind uses a double press of Esc to return the conversation and code to a safe point selected from the event log, with confirmation before work is undone ^[16].

The SDK allows a TypeScript program to start a local muse host, communicate with it over the Muse Session Protocol (MSP), stream session output, handle permission requests, and reload session state ^[16][18]. Meta published the SDK, protocol declarations, examples, and documentation under the MIT License. The repository explicitly says the real muse host binary is not included, so the SDK release should not be read as an open-source release of the Muse Code host or as a release of Muse Spark model weights ^[18].

Design

Meta describes Muse Code as a simple agent loop augmented by asynchronous background agents. These background agents stay alive for the whole session instead of being spawned per task, which Meta says avoids repeated information gathering and reduces latency on difficult multi-step work. They carry out next steps on their own and decide when to report back to the main agent ^[1].

Zuckerberg gave more detail in his launch posts: when a job is large enough, Muse Code fans out to sub-agents that work in parallel inside isolated worktrees, so the developer's working copy is never touched. In Meta's testing, the agent built six features for a game simultaneously without collisions ^[6][7].

The runtime is built around a local append-only event log that records every model call, tool run, approval, and edit. Meta says this makes the runtime replay-exact and restart-safe: after a crash, the agent resumes exactly where it stopped, which matters for tasks that run for hours ^[1][8]. Muse Code also ships with default skills: /plan turns a task into an approval-gated plan, /grill stress-tests that plan, and /goal keeps the agent working toward a stated objective ^[1][9].

Muse Spark 1.2

The model behind the agent is Muse Spark 1.2, a coding-focused update to Muse Spark 1.1. Meta says it scaled up training compute on coding tasks, widened training environment diversity, and trained extensively on long-horizon work such as whole-repository generation and large end-to-end projects, with planning, goal conditioning, and context compaction used to sustain progress ^[1]. Meta also describes a self-improvement loop: Muse Spark 1.1 generated challenging coding environments and instruction-following templates, then graded candidate solutions, producing training data for 1.2 ^[1].

As a capability demonstration, Meta had the model iteratively optimize GPU kernels over more than 1,000 tool calls and up to 24 hours per run, benchmarking KDA and MLA kernels written in Triton for NVIDIA Hopper GPUs against reference implementations. Models were barred from importing existing kernel libraries and had to implement the algorithms themselves ^[1][7].

Muse Spark 1.2 is available in Muse Code and through the Meta Model API "with expanded global access," a change from the US-only preview that accompanied 1.1 ^[1][11]. The model remains closed-weight. Asked on X whether Muse Spark would be open-sourced, Zuckerberg replied "I'll have more to share on that soon" ^[7]. Meta has not said whether the Muse Code client itself will be open source, and no public source repository for it had been identified as of August 7, 2026 ^[14].

Benchmark results

Meta published bar charts of its own evaluations rather than claiming state of the art. On every coding chart in the announcement, Muse Spark 1.2 with Muse Code trails Anthropic's Claude Opus 5 running in Claude Code and beats xAI's Grok 4.5 in Grok Build; against OpenAI's GPT-5.6 Terra in Codex it splits the results, leading on Terminal-Bench 2.1 and Meta Internal Coding Bench but trailing on DeepSWE 1.1 (59.3% versus 64.8%). The Register summed the showing up as "close to the best, but never the very best" ^[7]. All figures below are Meta's numbers from the August 5 announcement and should be read as vendor-reported ^[1][2].

The one chart Muse Spark 1.2 tops is MCP Atlas, a Scale AI benchmark of tool use across real Model Context Protocol servers ^[1][2]. GDPVal-AA v2 scores are Elo ratings produced by Artificial Analysis on professional deliverable tasks drawn from OpenAI's GDPval dataset ^[2].

Meta's accompanying methodology report explains how the numbers were produced. Terminal-Bench 2.1 (a Stanford and Laude Institute benchmark) was run on all 89 tasks of the official release, five attempts per task, each agent in an isolated cloud sandbox; DeepSWE v1.1 is Datacurve's 113-task software engineering benchmark spanning 91 repositories in five languages; Meta Internal Coding Bench is 440 tasks derived from Meta's internal pull requests. Each third-party model ran inside its own vendor's agent (Claude Code for Opus 5, Codex for GPT-5.6 Terra, Grok Build for Grok 4.5, Antigravity for Gemini 3.6 Flash), and Meta cautions that its setup may not show those models at their best ^[2].

The numbers are not official leaderboard entries. On the public Terminal-Bench 2.1 leaderboard at tbench.ai, which lists runs from many different agent harnesses, submitted through the benchmark's harbor framework and verified by the Terminal-Bench team, the top verified entry as of August 7, 2026 was Claude Code running Anthropic's Fable 5 model at 83.8%, and Muse Spark 1.2 did not yet appear; Muse Spark 1.1 sat at 76.2% with the mini-SWE-agent harness, matching Meta's chart ^[10]. Commenters on Hacker News and The Register also asked why Meta compared against OpenAI's GPT-5.6 Terra rather than the larger GPT-5.6 Sol on the coding charts ^[7][14].

Muse Spark 1.3

Muse Spark 1.3 became available in Muse Code and Meta Model API on September 2, 2026, with the install command unchanged (curl -fsSL https://dev.meta.ai/install.sh | bash) ^[22]. Meta's launch post says the model "was trained on more long-horizon coding tasks and shows improved usability in common engineering workflows," that relative to 1.2 "it takes fewer turns where not needed and is less verbose, while having a cleaner overall coding style," and that "in comparisons by Meta engineers, it proved to be significantly faster and more efficient, using ~20% fewer tool calls and ~25% fewer tokens" ^[22]. Those are Meta's internal comparisons; Winbuzzer noted that the company published neither the task count nor the accuracy pairing behind them ^[29].

Meta's benchmark chart for 1.3 lists Terminal-Bench 2.1 at 88.8 for Muse Spark 1.3 at max reasoning against 82.9 for 1.2 at xhigh, 88.8 for GPT-5.6 Sol (max) and 86.7 for Claude Opus 5 (max); the accompanying methodology says each task was run "with each model's native coding harnesses" inside Meta's evaluation framework, with the GPT-5.6 Sol figure taken from OpenAI's model card ^[26][27]. The same chart gives 75.4 on DeepSWE v1.1 and 59.4 on SWE-Atlas CodeBase QnA for 1.3 max, where the DeepSWE run used a mini-SWE-agent harness rather than Muse Code ^[26][27]. All of these are Meta-reported numbers.

The max reasoning mode was not part of the September 2 release. The launch post originally read "Previously available reasoning modes are available today with max reasoning coming shortly after we finish additional safety testing" ^[28]. On September 4 at 18:16 UTC Alexandr Wang posted "1/ we just publicly released Muse Spark 1.3 max!" and recommended trying it "even if you've already tried muse spark 1.3 high or muse spark 1.3 xhigh" ^[23], and the Meta for Developers account announced that "Muse Spark 1.3 with max reasoning is now available on Muse Code and Meta Model API" ^[24]. Meta then changed the launch post's availability line to that wording without adding an edit note ^[22][28].

Pricing did not change with the new model. Meta's developer pages list muse-spark-1.3 at $1.25 per million input tokens, $0.15 per million cached input tokens and $4.25 per million output tokens, and muse-spark-1.3-contributor at $0.10, $0.002 and $0.20, next to identical rows for the 1.2 variants; the 1 million token context window also carries over ^[25][26][29].

Pricing and availability

At the beta launch, Muse Code was metered through the Meta Model API, with billing tied to the same Meta developer account that hosted the API ^[5]. On August 31, Meta began rolling out three monthly subscriptions alongside the existing pay-as-you-go options ^[16][19].

Meta's product page also lists voice mode and web search for Everyday Usage, access to the latest models for High Usage, and early feature access and higher file-upload limits for Power Usage ^[20]. The 10-50 request figure is a published estimate rather than a fixed unit of coding work; Meta told The New Stack that the number varies with task complexity ^[19].

The usage-based API routes remain available. The Standard tier charges $1.25 per million input tokens and $4.25 per million output tokens. The Contributor tier charges $0.10 per million input tokens and $0.20 per million output tokens; in exchange, Meta may use prompts and completions to improve its products ^[9][20]. At launch, Wang called the Contributor route "more than 10 times cheaper than even the pay-as-you-go tier" ^[5].

Token pricing and launch rate limits are from MacRumors and CNBC, August 5, 2026 ^[5][9]. Meta's current product page continues to list both token-priced routes ^[20]; the historic rate limits remain sourced to the launch coverage and Meta's API documentation ^[13].

Wang also said Meta was "starting to accept requests for zero-data retention" for enterprises that do not want any developer data kept ^[5]. Beyond Meta's own API, Muse Spark 1.2 is available through OpenRouter ^[5][7]. The model offers a 1 million token context window, a capacity introduced with Muse Spark 1.1 ^[7][11].

Competitive landscape

Muse Code enters a crowded field of terminal agents: Claude Code, OpenAI Codex, Gemini CLI, and editor-centric tools such as Cursor. Wang told CNBC that Meta is differentiating on price rather than capability against Anthropic and OpenAI ^[5], and told the Wall Street Journal the tool "can be an incredibly good option, especially from a cost perspective" for many workflows ^[6]. The Contributor tier's discount, roughly a twelfth of standard input pricing, is the sharpest expression of that strategy, though it revives a familiar Meta trade-off: MacRumors headlined its coverage "Meta's New Mac Coding Agent Costs Up to 20x Less If You Let Meta Train on Your Data" ^[9].

The launch also fits a broader monetization push. CNBC noted that Meta, which still earns 98% of its revenue from advertising, is trying to build AI revenue while spending heavily on data centers, and that its shares fell the prior week on a light revenue forecast and shrinking free cash flow ^[5].

Reception

Early reviews treated Muse Code as a credible but not category-leading entry. The Register judged that Muse Spark 1.2 "performed honorably" against Opus 5, GPT-5.6 Terra, Grok 4.5, and Gemini 3.6 in tightly clustered scores, and noted researcher Hongyu Ren of Meta Superintelligence Labs boasting "TBH it's a good harness" ^[7]. On Hacker News, commenters welcomed the expanded global API access and the cheap Contributor pricing, comparing it to DeepSeek-level rates, but criticized the required Meta account sign-in, pressed Meta to publish cost and latency alongside benchmark scores, and repeatedly asked whether the client or the model weights would be opened ^[14]. The closed-weight release continues the break from Meta's Llama-era openness, which The Register called out directly ^[7].

See also

• Muse Spark
• Meta Superintelligence Labs
• Claude Code
• OpenAI Codex
• Gemini CLI
• Terminal-Bench
• SWE-bench
• Meta AI

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18Meta Superintelligence Labs. "Introducing Muse Code and Muse Spark 1.2." Meta AI Research Blog. August 5, 2026. research.meta.ai/...g-muse-code-and-muse-spark-1-2
2. ^1 ^2 ^3 ^4Meta Superintelligence Labs. "Muse Spark 1.2 & Muse Code: Evaluation Methodology." August 2026. research.meta.ai/...muse-spark-1-2-methodology
3. ^1 ^2 ^3AI at Meta (@AIatMeta). "Introducing Muse Code (beta), a terminal coding agent built for long-horizon software engineering..." X. August 5, 2026. x.com/...2085084709277565213
4. ^Mark Zuckerberg (@finkd). "Releasing Muse Code in beta today..." X. August 5, 2026. x.com/...2085080750034940201
5. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10Jonathan Vanian. "Meta debuts first AI coding agent to take on Anthropic and OpenAI." CNBC. August 5, 2026. cnbc.com/...-code-to-take-on-anthropic-and-openai-
6. ^1 ^2 ^3Lucas Ropek. "Meta launches Muse Code, an AI agent for large code bases." TechCrunch. August 5, 2026. techcrunch.com/...an-ai-agent-for-large-code-bases
7. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11Joab Jackson. "Meta wants to get inside your terminal with its new coding agent." The Register. August 6, 2026. theregister.com/...5283717
8. ^1 ^2 ^3Zac Hall. "Meta launches Muse Code AI coding agent for macOS and Linux." 9to5Mac. August 5, 2026. 9to5mac.com/...ai-coding-agent-for-macos-and-linux
9. ^1 ^2 ^3 ^4 ^5 ^6Juli Clover. "Meta's New Mac Coding Agent Costs Up to 20x Less If You Let Meta Train on Your Data." MacRumors. August 5, 2026. macrumors.com/...meta-muse-code-for-mac
10. ^Terminal-Bench. "terminal-bench@2.1 Leaderboard." tbench.ai. Accessed August 7, 2026. tbench.ai/...2.1
11. ^1 ^2 ^3Meta Superintelligence Labs. "Introducing Muse Spark 1.1." Meta AI Research Blog. July 9, 2026. research.meta.ai/...cing-muse-spark-meta-model-api
12. ^1 ^2Lucas Ropek. "Meta enters the crowded AI coding battle with Muse Spark 1.1." TechCrunch. July 9, 2026. techcrunch.com/...oding-battle-with-muse-spark-1-1
13. ^Meta. "Model API: Pricing and rate limits." dev.meta.ai. Accessed August 7, 2026. dev.meta.ai/...pricing-rate-limits
14. ^1 ^2 ^3 ^4Hacker News. "Muse Code and Muse Spark 1.2" (discussion thread). August 5, 2026. news.ycombinator.com/item
15. ^Meta Superintelligence Labs. "Introducing Muse Spark: Scaling Towards Personal Superintelligence." Meta AI Blog. April 8, 2026. ai.meta.com/...introducing-muse-spark-msl
16. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11Claire Zhou. "Muse Code: New plans and features." Meta for Developers. August 31, 2026. developer.meta.com/...-code-new-plans-and-features
17. ^1 ^2 ^3Alexandr Wang (@alexandr_wang). "muse code is out of beta..." X. August 31, 2026. x.com/...2094502557129543774
18. ^1 ^2 ^3 ^4Meta Models. "Muse Code SDK." GitHub. Accessed September 2, 2026. github.com/...muse-code-sdk
19. ^1 ^2 ^3 ^4 ^5 ^6Paul Sawers. "Meta's Claude Code rival exits beta with three new subscription tiers, and it's pushing hard on price." The New Stack. September 1, 2026. thenewstack.io/muse-code-sdk-pricing
20. ^1 ^2 ^3 ^4 ^5 ^6 ^7Meta. "Muse Code." Meta for Developers. Accessed September 2, 2026. developer.meta.com/...muse-code
21. ^Meta Models. "Muse Code SDK changelog." GitHub. Accessed September 2, 2026. github.com/...CHANGELOG.md
22. ^1 ^2 ^3 ^4 ^5Meta Superintelligence Labs. "Introducing Muse Spark 1.3." Meta AI Research Blog. September 2, 2026; availability line updated in place by September 5, 2026. research.meta.ai/...introducing-muse-spark-1-3
23. ^1 ^2 ^3Alexandr Wang (@alexandr_wang). "1/ we just publicly released Muse Spark 1.3 max!..." X. September 4, 2026. x.com/...2095938990197329935
24. ^1 ^2 ^3Meta for Developers (@MetaforDevs). "Muse Spark 1.3 with max reasoning is now available on Muse Code and Meta Model API..." X. September 4, 2026. x.com/...2095939388559720720
25. ^1 ^2Meta. "Muse Code." Meta for Developers, as archived by the Internet Archive on September 4, 2026 (17:25 UTC). web.archive.org/...muse-code
26. ^1 ^2 ^3Meta. "Muse Spark 1.3." Meta for Developers, as archived by the Internet Archive on September 3, 2026. web.archive.org/...muse-spark
27. ^1 ^2Meta Superintelligence Labs. "Muse Spark 1.3 Evaluation Methodology." September 2026. research.meta.ai/...timodal-evaluation-methodology
28. ^1 ^2Meta Superintelligence Labs. "Introducing Muse Spark 1.3," as archived by the Internet Archive on September 2, 2026 (19:37 UTC). web.archive.org/...introducing-muse-spark-1-3
29. ^1 ^2Markus Kasanmascheff. "Meta Releases Muse Spark 1.3 Model for Longer Tool-Based Work." Winbuzzer. September 4, 2026. winbuzzer.com/...del-longer-tool-based-work-xcxwbn