# Muse Code

> Source: https://aiwiki.ai/wiki/muse_code
> Summary: 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 , then said it had left beta on August 31 .
> Updated: 2026-10-01
> Fact-checked: 2026-10-01
> Categories: AI Agents, AI Code Generation, AI Tools & Products, Developer Tools, Meta AI
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Muse Code." aiwiki.ai, 1 Oct 2026. https://aiwiki.ai/wiki/muse_code
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

| Field | Value |
| --- | --- |
| Developer | [Meta Superintelligence Labs](https://aiwiki.ai/wiki/meta_superintelligence_labs) (Meta Platforms) |
| Type | Terminal (command-line) [coding agent](https://aiwiki.ai/wiki/coding_agent) |
| Initial release | August 5, 2026 (beta) [1] |
| Beta ended | August 31, 2026 [16][17] |
| Underlying model | [Muse Spark](https://aiwiki.ai/wiki/muse_spark) 1.2 at launch [1]; Muse Spark 1.3 from September 2, 2026, with max reasoning from September 4 [22][23][24] |
| Platforms | macOS, Linux, Windows (Windows from September 16, 2026) [1][8][30][34] |
| Installation | `curl -fsSL https://dev.meta.ai/install.sh \| bash` (macOS, Linux); `irm https://dev.meta.ai/install.ps1 \| iex` (Windows) [1][30][34] |
| Pricing | Monthly plans from $5 plus usage-based Meta Model API tiers [16][19][20] |
| License | Muse Code proprietary; SDK MIT [7][18] |
| Latest release | 1.4.2, the newest entry in the published changelog as of October 1, 2026 [32][40] |
| Website | developer.meta.com/ai/products/muse-code/ |

**Muse Code** is a terminal coding agent developed by [Meta Superintelligence Labs](https://aiwiki.ai/wiki/meta_superintelligence_labs). Meta introduced it in beta on August 5, 2026 alongside the [Muse Spark](https://aiwiki.ai/wiki/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](https://aiwiki.ai/wiki/claude_code), OpenAI's [Codex](https://aiwiki.ai/wiki/openai_codex), and Google's [Gemini CLI](https://aiwiki.ai/wiki/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].

Meta kept shipping through September 2026. A native Windows build arrived on September 16 [30]; at [Meta Connect 2026](https://aiwiki.ai/wiki/meta_connect_2026) the company said the [Meta Model API](https://aiwiki.ai/wiki/meta_model_api) behind the agent had become generally available worldwide [38]; and on September 29 the Meta for Developers account announced Monitor, a background watcher that wakes the agent when something it is watching changes [31]. The newest entry in Meta's published Muse Code changelog as of October 1, 2026 is version 1.4.2 [32].

## Background

Meta Superintelligence Labs, the unit [Mark Zuckerberg](https://aiwiki.ai/wiki/mark_zuckerberg) formed in June 2025 around former [Scale AI](https://aiwiki.ai/wiki/scale_ai) chief executive [Alexandr Wang](https://aiwiki.ai/wiki/alexandr_wang), replaced the open-weight [Llama](https://aiwiki.ai/wiki/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](https://aiwiki.ai/wiki/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]. Meta's documentation as of October 1, 2026 states the requirement differently: Workflows need a build that includes the workflow script engine plus access to a staged rollout, and the public Apple silicon macOS package does not ship that engine, so Workflows are unavailable in it [35]. A workflow may call up to 1,000 child tasks over its lifetime, with the number running at once derived from the host CPU count and capped at 16 [35]. 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]. It is published on npm as `@muse-code/sdk`, needs Node.js 20 or newer, and ships with no runtime dependencies; the repository README calls the SDK and its documentation a Developer Preview with no stability promise [18][40]. The changelog's 1.4.0 entry tells users to install the Python SDK from PyPI, where `muse-code-sdk` first appeared as version 1.3.1 on September 21, 2026 alongside a wire-types companion package `muse-code-msp` [32][41]. 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].

## Windows support

Muse Code reached Windows on September 16, 2026, when the Meta for Developers account said the agent was "now available natively on Windows - no WSL required" and described it as "PowerShell-fluent, sandboxed by default, native x64 and ARM64" with "zero-admin install," installed with `irm https://dev.meta.ai/install.ps1 | iex` [30]. Meta's documentation says the agent runs on macOS, Linux and Windows from one codebase, and lists the differences on Windows: agent-run commands use PowerShell syntax and Windows command-line tools, the sandbox can raise a Windows User Account Control prompt the first time it initializes, and voice input and session messaging are not available; on voice input the changelog disagrees with the documentation, its 1.4.0 entry saying voice input "is enabled by default on Windows x64 and ARM64" [32][34][43]. The September announcement listed inter-session messaging as coming to Windows later [30]. Meta repeated the Windows availability in its Meta Connect 2026 recap, which described Muse Code as "now out of beta and available on Windows" and "multi-agent by default, allowing parallel workers to write and test code in isolated workspaces" [38].

## Monitor

On September 29, 2026 the Meta for Developers account announced Monitor. The post says Muse Code "watches CI, builds, and logs in the background and wakes the agent when there's something to act on," adding "No model calls while it waits," and suggests the prompt "Open the PR, watch the checks, and fix anything that fails" [31]. The attached screenshot shows a status row reading "Watching  Watch PR 43767 CI, comments, conflicts" marked `persistent`, with a hint that the down arrow lists active monitors, the same watch repeated as a child row under the main agent with the state `watching`, and a separate composer hint that `/loop 10m <prompt>` "schedules a recurring prompt" [31].

Meta's published documentation covers monitors mainly through the surfaces that manage them and through the release notes. The changelog's 1.4.0 entry lists "The `monitor` tool is now available by default" among its improvements, and records that `/stop` and the Tasks stop actions "now use the same durable Work Stop path for workflows, background commands, and monitors" [32]. Monitors are background work in the same sense as long shell commands: the `/tasks` drawer manages them, `/stop` cancels applicable background tasks and terminals rather than the current turn, and an agent can stop a monitor it started with the `work_stop` tool [32][33].

Other 1.4.x notes fill in the behavior. A monitor can be persistent, and it produces wake turns that enter the model's context; forking or rewinding a session carries those wake bodies into the branch [32]. The changelog refers to a persistent `ws` monitor, which is a watch held open on a WebSocket connection, and to monitor commands that contain web addresses [32]. Stopping a watch produces one completion with its reason available in the expanded details, and a monitor armed while a `muse exec` run is finishing is stopped with a receipt instead of failing to start [32]. Release 1.4.2 added a full-screen monitor view that "shows the command or script the watch runs, with secret values masked," and changed goal tracking so that while a `/goal` is active the runtime can pause its automatic follow-up turns for up to an hour while it waits on a background command or monitor rather than checking repeatedly [32].

In a feature request filed on Meta's SDK issue tracker on September 25, 2026, an outside developer reported that `monitor` was among several built-in tools that never reached `muse serve` sessions in Muse Code 1.3.0 on Windows, and summarized the tool's own description in the binary as watching "a command or a `ws://` / `wss://` stream" to wake the agent [42]. That report is a third-party observation rather than Meta documentation.

Scheduled prompts are a separate mechanism from monitors. `/loop` registers a recurring prompt as a cron job, accepting either a shorthand interval such as `5m`, `1h` or `2d` or a five-field cron expression in local time, and defaults to every ten minutes; an occurrence that lands during an active run is skipped rather than queued or replayed, and the scheduled text can be a plain instruction or a skill invocation passed through verbatim [33]. Meta's documentation page and its changelog disagree on how long a loop lives: the page says recurring jobs auto-expire after seven days, while the 1.4.0 release notes say a new `/loop` job "now runs until you delete it instead of stopping on its own after about a week" and tell users to recreate loops made by earlier versions [32][33].

Anthropic documents a comparable pairing for [Claude Code](https://aiwiki.ai/wiki/claude_code): its scheduled-prompts page says that in a session where its own Monitor tool is available, the model may use it for a dynamic `/loop` schedule because "Monitor runs a background script and streams each output line back, which avoids polling altogether and is often more token-efficient and responsive than re-running a prompt on an interval" [44].

## Releases since the launch

Meta publishes a changelog for Muse Code on its documentation site. The entries carry version numbers but no dates; the matching SDK packages give approximate timing, with `@muse-code/sdk` 1.3.0 published on npm on September 18, 2026 and 1.4.2 on September 30, 2026 [32][40].

| Release | Selected additions |
| --- | --- |
| 1.4.0 | `monitor` tool on by default; Python SDK on PyPI; `/resume-claude` and `/resume-codex` to continue local [Claude Code](https://aiwiki.ai/wiki/claude_code) or [Codex](https://aiwiki.ai/wiki/openai_codex) sessions; remaining subscription usage readable over the session protocol without spending a prompt; saved MCP tool and network approvals; `/hud` status-bar picker; `muse exec --output-schema`; hook `onFailure` fallbacks; Vim composer editing [32] |
| 1.4.1 | Reasoning-effort tiers exposed to protocol clients; fixes to long-session live views and subagent recovery [32] |
| 1.4.2 | Full-screen monitor view with masked secrets; several workspace roots per session; `muse voice transcribe <file>`; session list filtering; `/delete` confirmation for a session and its children; goal tracking that waits on background work instead of polling [32] |

## 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]. Meta's documentation adds two further details about the agent tree. One tree runs eight agents at once by default, including the root, settable from 1 to 64 in `settings.json`, and cancellation is cooperative, so a cancelled child that has not reached a checkpoint keeps running [45]. Separately, four background observer agents run alongside the main session, one each for memory recall, skill recall, goal tracking and verification; each proposes a short advisory that a reconciler may or may not pass to the main agent, and because each makes its own model calls the default set adds token usage [45].

## 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](https://aiwiki.ai/wiki/triton) for [NVIDIA](https://aiwiki.ai/wiki/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](https://aiwiki.ai/wiki/claude_opus_5) running in Claude Code and beats xAI's [Grok](https://aiwiki.ai/wiki/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].

| Benchmark | Muse Spark 1.2 (Muse Code) | Best competitor in Meta's chart | Muse Spark 1.1 |
| --- | --- | --- | --- |
| [Terminal-Bench](https://aiwiki.ai/wiki/terminal_bench) 2.1 | 82.9% | Claude Opus 5 + Claude Code, 86.7% | 76.2% |
| DeepSWE 1.1 | 59.3% | Claude Opus 5, 65.0% (GPT-5.6 Terra 64.8%) | 53.0% |
| Meta Internal Coding Bench | 70.6% | Claude Opus 5, 79.4% | 68.3% |
| GDPVal-AA v2 (Elo) | 1631 | Claude Opus 5, 1852 | 1371 |
| MCP Atlas | 90.3% | Muse Spark 1.2 leads; next is 1.1 at 88.1% | 88.1% |

The one chart Muse Spark 1.2 tops is MCP Atlas, a Scale AI benchmark of tool use across real [Model Context Protocol](https://aiwiki.ai/wiki/model_context_protocol) servers [1][2]. GDPVal-AA v2 scores are Elo ratings produced by [Artificial Analysis](https://aiwiki.ai/wiki/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]. As of October 1, 2026 that leaderboard was topped by OpenAI's [GPT-6 Astra](https://aiwiki.ai/wiki/gpt_6_astra) (high) in Codex at 87.4% ± 1.8%, with Fable 5 (xhigh) in Claude Code second at 83.8% ± 2.3%; Muse Spark 1.1 (xhigh) with mini-SWE-agent held ninth place at 76.2% ± 2.4%, and no Muse Spark 1.2 or 1.3 run had been added [39]. 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.

Meta's Muse Spark product page carried a wider version of the same comparison as of October 1, 2026, setting Muse Spark 1.3 at `max` against Muse Spark 1.2 at `xhigh`, OpenAI's GPT-5.6 Sol at `max` and Anthropic's Claude Opus 5 at `max` across eleven benchmarks. Muse Spark 1.3 posted the highest reported figures on the two long-context MRCR splits (98.5 on the 256K to 512K split and 98.1 on the 512K to 1M split, against 91.5 and 73.8 for GPT-5.6 Sol, with no figure listed for Opus 5), and trailed on knowledge work, where its GDPVal-AA v2 Elo of 1754 sat below Opus 5 at 1824 and above GPT-5.6 Sol at 1710 [47].

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].

| Subscription | Monthly price | Published allowance (September 2026) | Published allowance (as of October 1, 2026) |
| --- | --- | --- | --- |
| Everyday Usage | $5 | 10-50 requests every five hours, including image and video uploads | 10-50 prompts every five hours, including image and video uploads |
| High Usage | $15 | Three times the usage of Everyday Usage | Five times the usage of Everyday Usage |
| Power Usage | $50 | Ten times the usage of Everyday Usage | Twenty times the usage of Everyday Usage |

The multipliers changed during September. The product page still read "3x" and "10x" on September 4, 2026, and by October 1 both Meta's product page and its subscriptions documentation read "5x" and "20x" [25][36][37]. Meta also moved the "latest Muse models" line from High Usage to Everyday Usage, which as of October 1, 2026 lists access to the latest Muse models, the 10-50 prompt allowance, voice mode and web search; High Usage adds more prompts with the latest models and more multimodal inputs, and Power Usage adds expanded prompts, early access to new features and higher file uploads [36][37]. The 10-50 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]. A subscription works only through the Muse Code command-line tool while signed in with a Meta Model API account, and any additional API keys created under the same account are billed pay-as-you-go [37].

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].

| Tier | Input (per 1M tokens) | Cached input (per 1M tokens) | Output (per 1M tokens) | Data use | Rate limits (as of October 1, 2026) |
| --- | --- | --- | --- | --- | --- |
| Standard | $1.25 | $0.15 | $4.25 | Not used for training | 3,000 requests/min, 4,000,000 tokens/min |
| Contributor | $0.10 | $0.002 | $0.20 | Prompts and completions may train Meta models | 100 requests/min, 3,000,000 tokens/min |

Token pricing is from MacRumors and CNBC, August 5, 2026, and is unchanged on Meta's product, model and pricing pages as of October 1, 2026 [5][9][36][46]. Launch coverage and Meta's API documentation as read on August 7, 2026 put the Contributor tier at 60 requests and 2.1 million tokens per minute [13]; the documentation now lists 100 requests and 3,000,000 tokens per minute for that tier, with the Standard tier unchanged [46]. The Standard rates cover `muse-spark-1.3`, `muse-spark-1.2` and `muse-spark-1.1`, and the Contributor rates cover the 1.3 and 1.2 contributor variants; Meta says there is no long-context premium, and web search grounding costs $2.50 per 1,000 queries on top of token cost [46].

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](https://aiwiki.ai/wiki/openrouter) [5][7]. The model offers a 1 million token context window, a capacity introduced with Muse Spark 1.1 [7][11].

Meta's Meta Connect 2026 recap, published September 24, 2026, said the Meta Model API had become "generally available globally, with support and compliance for enterprise workloads through our first-party API and third-party partners," and that Muse Spark 1.3 was available through Oracle Cloud AI Platform with [Google Cloud](https://aiwiki.ai/wiki/google_cloud) availability in private preview [38]. The same post announced a Meta Global AI Developer Hackathon, a ten-day virtual event with free Model API credits, workshops run by Meta Superintelligence Labs staff, separate tracks for independent developers and startup teams, and a $1 million cash prize pool, open to projects built on the Meta Model API, Muse Code or Meta's open-weight models [38]. Meta's documentation says the default model on first run is `muse-spark-1.2`, with other models selectable in the session [34].

## Competitive landscape

Muse Code enters a crowded field of terminal agents: Claude Code, OpenAI Codex, Gemini CLI, and editor-centric tools such as [Cursor](https://aiwiki.ai/wiki/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](https://aiwiki.ai/wiki/muse_spark)
- [Meta Superintelligence Labs](https://aiwiki.ai/wiki/meta_superintelligence_labs)
- [Claude Code](https://aiwiki.ai/wiki/claude_code)
- [OpenAI Codex](https://aiwiki.ai/wiki/openai_codex)
- [Gemini CLI](https://aiwiki.ai/wiki/gemini_cli)
- [Terminal-Bench](https://aiwiki.ai/wiki/terminal_bench)
- [SWE-bench](https://aiwiki.ai/wiki/swe_bench)
- [Meta AI](https://aiwiki.ai/wiki/meta_ai)
- [Meta Model API](https://aiwiki.ai/wiki/meta_model_api)
- [Meta Connect 2026](https://aiwiki.ai/wiki/meta_connect_2026)

## References

1. Meta Superintelligence Labs. "Introducing Muse Code and Muse Spark 1.2." Meta AI Research Blog. August 5, 2026. https://research.meta.ai/blog/introducing-muse-code-and-muse-spark-1-2
2. Meta Superintelligence Labs. "Muse Spark 1.2 & Muse Code: Evaluation Methodology." August 2026. https://research.meta.ai/static/muse-spark-1-2-methodology
3. AI at Meta (@AIatMeta). "Introducing Muse Code (beta), a terminal coding agent built for long-horizon software engineering..." X. August 5, 2026. https://x.com/AIatMeta/status/2085084709277565213
4. Mark Zuckerberg (@finkd). "Releasing Muse Code in beta today..." X. August 5, 2026. https://x.com/finkd/status/2085080750034940201
5. Jonathan Vanian. "Meta debuts first AI coding agent to take on Anthropic and OpenAI." CNBC. August 5, 2026. https://www.cnbc.com/2026/08/05/meta-debuts-muse-code-to-take-on-anthropic-and-openai-.html
6. Lucas Ropek. "Meta launches Muse Code, an AI agent for large code bases." TechCrunch. August 5, 2026. https://techcrunch.com/2026/08/05/meta-launches-muse-code-an-ai-agent-for-large-code-bases/
7. Joab Jackson. "Meta wants to get inside your terminal with its new coding agent." The Register. August 6, 2026. https://www.theregister.com/ai-and-ml/2026/08/06/meta-wants-to-get-inside-your-terminal-with-its-new-coding-agent/5283717
8. Zac Hall. "Meta launches Muse Code AI coding agent for macOS and Linux." 9to5Mac. August 5, 2026. https://9to5mac.com/2026/08/05/meta-launches-muse-code-ai-coding-agent-for-macos-and-linux/
9. Juli Clover. "Meta's New Mac Coding Agent Costs Up to 20x Less If You Let Meta Train on Your Data." MacRumors. August 5, 2026. https://www.macrumors.com/2026/08/05/meta-muse-code-for-mac/
10. Terminal-Bench. "terminal-bench@2.1 Leaderboard." tbench.ai. Accessed August 7, 2026. https://www.tbench.ai/leaderboard/terminal-bench/2.1
11. Meta Superintelligence Labs. "Introducing Muse Spark 1.1." Meta AI Research Blog. July 9, 2026. https://research.meta.ai/blog/introducing-muse-spark-meta-model-api
12. Lucas Ropek. "Meta enters the crowded AI coding battle with Muse Spark 1.1." TechCrunch. July 9, 2026. https://techcrunch.com/2026/07/09/meta-enters-the-crowded-ai-coding-battle-with-muse-spark-1-1/
13. Meta. "Model API: Pricing and rate limits." dev.meta.ai. Accessed August 7, 2026. https://dev.meta.ai/docs/pricing-rate-limits
14. Hacker News. "Muse Code and Muse Spark 1.2" (discussion thread). August 5, 2026. https://news.ycombinator.com/item?id=49187575
15. Meta Superintelligence Labs. "Introducing Muse Spark: Scaling Towards Personal Superintelligence." Meta AI Blog. April 8, 2026. https://ai.meta.com/blog/introducing-muse-spark-msl/
16. Claire Zhou. "Muse Code: New plans and features." Meta for Developers. August 31, 2026. https://developer.meta.com/ai/resources/blog/muse-code-new-plans-and-features/
17. Alexandr Wang (@alexandr_wang). "muse code is out of beta..." X. August 31, 2026. https://x.com/alexandr_wang/status/2094502557129543774
18. Meta Models. "Muse Code SDK." GitHub. Accessed September 2, 2026. https://github.com/meta-models/muse-code-sdk
19. Paul 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. https://thenewstack.io/muse-code-sdk-pricing/
20. Meta. "Muse Code." Meta for Developers. Accessed September 2, 2026. https://developer.meta.com/ai/products/muse-code/
21. Meta Models. "Muse Code SDK changelog." GitHub. Accessed September 2, 2026. https://github.com/meta-models/muse-code-sdk/blob/main/CHANGELOG.md
22. Meta Superintelligence Labs. "Introducing Muse Spark 1.3." Meta AI Research Blog. September 2, 2026; availability line updated in place by September 5, 2026. https://research.meta.ai/blog/introducing-muse-spark-1-3
23. Alexandr Wang (@alexandr_wang). "1/ we just publicly released Muse Spark 1.3 max!..." X. September 4, 2026. https://x.com/alexandr_wang/status/2095938990197329935
24. Meta for Developers (@MetaforDevs). "Muse Spark 1.3 with max reasoning is now available on Muse Code and Meta Model API..." X. September 4, 2026. https://x.com/MetaforDevs/status/2095939388559720720
25. Meta. "Muse Code." Meta for Developers, as archived by the Internet Archive on September 4, 2026 (17:25 UTC). https://web.archive.org/web/20260904172529/https://developer.meta.com/ai/products/muse-code/
26. Meta. "Muse Spark 1.3." Meta for Developers, as archived by the Internet Archive on September 3, 2026. https://web.archive.org/web/20260903083136/https://developer.meta.com/ai/models/muse-spark/
27. Meta Superintelligence Labs. "Muse Spark 1.3 Evaluation Methodology." September 2026. https://research.meta.ai/static/muse-spark-1-3-multimodal-evaluation-methodology
28. Meta Superintelligence Labs. "Introducing Muse Spark 1.3," as archived by the Internet Archive on September 2, 2026 (19:37 UTC). https://web.archive.org/web/20260902193756/https://research.meta.ai/blog/introducing-muse-spark-1-3
29. Markus Kasanmascheff. "Meta Releases Muse Spark 1.3 Model for Longer Tool-Based Work." Winbuzzer. September 4, 2026. https://winbuzzer.com/2026/09/04/meta-releases-muse-spark-1-3-model-longer-tool-based-work-xcxwbn/
30. Meta for Developers (@MetaforDevs). "Muse Code from Meta is now available natively on Windows - no WSL required..." X. September 16, 2026. https://x.com/MetaforDevs/status/2100268678583566691
31. Meta for Developers (@MetaforDevs). "New in Muse Code: Monitor..." X. September 29, 2026. https://x.com/MetaforDevs/status/2104986078969319445
32. Meta. "Changelog." Muse Code documentation. Accessed October 1, 2026. https://dev.meta.ai/docs/muse-code/changelog
33. Meta. "Working with the agent." Muse Code documentation. Accessed October 1, 2026. https://dev.meta.ai/docs/muse-code/interactive
34. Meta. "Muse Code." Meta Model API documentation. Accessed October 1, 2026. https://dev.meta.ai/docs/muse-code
35. Meta. "Run multi-agent workflows." Muse Code documentation. Accessed October 1, 2026. https://dev.meta.ai/docs/muse-code/workflows
36. Meta. "Muse Code." Meta for Developers. Accessed October 1, 2026. https://developer.meta.com/ai/products/muse-code/
37. Meta. "Subscriptions." Muse Code documentation. Accessed October 1, 2026. https://dev.meta.ai/docs/muse-code/subscriptions
38. Meta. "Meta Connect 2026: The end-to-end recap." Meta for Developers blog. September 24, 2026. https://developers.meta.com/blog/meta-connect-recap/
39. Terminal-Bench. "terminal-bench@2.1 Leaderboard." tbench.ai. Accessed October 1, 2026. https://www.tbench.ai/leaderboard/terminal-bench/2.1
40. npm registry metadata for `@muse-code/sdk`. Accessed October 1, 2026. https://registry.npmjs.org/@muse-code%2Fsdk
41. PyPI release metadata for `muse-code-sdk`. Accessed October 1, 2026. https://pypi.org/pypi/muse-code-sdk/json
42. RandyNorthrup. "[Feature]: Let MSP clients see a session's tool set and turn on the tools `muse serve` leaves off (image_generation, web_fetch, artifact, monitor, code mode, tool_search), with their billing documented." Issue #40, meta-models/muse-code-sdk, GitHub. September 25, 2026. https://github.com/meta-models/muse-code-sdk/issues/40
43. Meta. "Coordinate sessions with messages." Muse Code documentation. Accessed October 1, 2026. https://dev.meta.ai/docs/muse-code/session-messaging
44. Anthropic. "Run prompts on a schedule." Claude Code Docs. Accessed October 1, 2026. https://code.claude.com/docs/en/scheduled-tasks
45. Meta. "Extending and automating." Muse Code documentation. Accessed October 1, 2026. https://dev.meta.ai/docs/muse-code/extending
46. Meta. "Pricing and rate limits." Meta Model API documentation. Accessed October 1, 2026. https://dev.meta.ai/docs/pricing-rate-limits
47. Meta. "Muse Spark 1.3." Meta for Developers. Accessed October 1, 2026. https://developer.meta.com/ai/models/muse-spark/