{"slug": "meta-muse-code-beta-what-developers-actually-get", "title": "Meta Muse Code Beta: What Developers Actually Get", "summary": "Meta launched Muse Code, a terminal coding agent that defaults to training on user code unless switched to the Standard tier, which costs $1.25 per million input tokens and $4.25 per million output tokens. The beta runs on Muse Spark 1.2 and features parallel sub-agents in isolated worktrees, a crash-safe event log, and persistent background agents. Benchmarks show Muse Spark 1.2 scores 82.9% on Terminal-Bench 2.1, 59.3% on DeepSWE v1.1, and 70.6% on Meta Internal Coding Bench, trailing Claude Opus 5.", "body_md": "Meta shipped a terminal coding agent this week, and there is one thing you need to know before you install it: it defaults to training on your code. [Muse Code](https://techcrunch.com/2026/08/05/meta-launches-muse-code-an-ai-agent-for-large-code-bases/) (beta) runs on Muse Spark 1.2, costs $1.25 per million input tokens on the safe tier, and brings three architecture moves that no other terminal agent has shipped yet. Whether it earns a spot in your workflow is a separate question from whether it is impressive. For the right workloads, it is both.\n\n## Install in One Command, Then Immediately Change One Setting\n\nInstallation is a single shell command on macOS and Linux:\n\n```\ncurl -fsSL https://dev.meta.ai/install.sh | bash\n```\n\nAfter that, open your settings and switch tiers. Muse Code defaults to the Contributor tier, which charges $0.10 per million input tokens and $0.20 per million output tokens. That sounds like a bargain until you read the terms: Meta trains on your inputs and outputs, and those rights cannot be revoked once your code enters the model weights. There is also a 100-request-per-minute cap on the contributor tier, which makes the parallel agent features — the ones actually worth having — unusable at scale.\n\nThe Standard tier charges $1.25 per million input tokens and $4.25 per million output tokens. That is still 8 to 12 times cheaper than Claude Opus 5 or GPT-5.6 Sol on token economics alone. Switch to it before you run anything on code you care about.\n\n| Tier | Input (per 1M tokens) | Output (per 1M tokens) | Rate Limit | Your Code |\n|---|---|---|---|---|\n| Standard | $1.25 | $4.25 | 3,000 RPM | Not used for training |\n| Contributor | $0.10 | $0.20 | 100 RPM | Meta trains on it |\n\n## The Three Features That Actually Differentiate It\n\nMuse Code does the standard terminal agent things — plans tasks, writes code, runs tests, edits files across a repo — but its architecture has three moves that no current competitor has shipped as a package.\n\n### Parallel Sub-Agents in Isolated Worktrees\n\nWhen a task is complex enough, Muse Code fans out to child agents running in separate git worktrees branched from the lead’s commit. Your primary working directory stays untouched. Concurrency scales with your core count, clamped between 2 and 16. Zuckerberg demoed building six game features simultaneously with zero file collisions. That is the kind of workflow acceleration that batch refactors and large-scale test generation actually need.\n\n### Crash-Safe Event Log\n\nEvery model call, tool run, approval, and file edit is appended to a local write-ahead log before it executes. If the agent crashes mid-task — and terminal agents do crash — it resumes exactly where it stopped. No re-explaining context, no lost progress. This solves a real operational problem that anyone who has watched a long-running agent die at 80% completion will immediately recognize.\n\n### Persistent Background Agents\n\nMost agents spawn fresh workers per subtask. Muse Code keeps four background agents alive throughout the entire session: memory recall, skill recall, goal tracking, and an optional verification agent. The session-persistent design means context is gathered once rather than re-fetched on every subtask, which cuts latency on long multi-step work. The trade-off is that each background agent makes independent model calls, so your token bill includes overhead you do not explicitly request.\n\n## Benchmarks: Honest Numbers\n\n[Meta published results](https://research.meta.ai/blog/introducing-muse-code-and-muse-spark-1-2) from its own eval harness comparing Muse Spark 1.2 against Claude Opus 5, GPT-5.6 Terra, Grok 4.5, and Kimi K3. Muse Spark 1.2 places second to Claude Opus 5 on all three benchmarks.\n\n| Benchmark | Claude Opus 5 | Muse Spark 1.2 |\n|---|---|---|\n| Terminal-Bench 2.1 | 86.7% | 82.9% |\n| DeepSWE v1.1 | 65.0% | 59.3% |\n| Meta Internal Coding Bench | 79.4% | 70.6% |\n\nA few things to hold onto: Meta compared against GPT-5.6 Terra rather than the flagship Sol. [Independent testers found](https://kingy.ai/blog/muse-code-muse-spark-1-2-benchmarks-verified/) that extracting Muse Code’s system prompt and running it on a competing model produced results that were faster and cheaper — which suggests some of the apparent efficiency gains come from prompt engineering, not raw model capability. There are also no independent leaderboard entries for Muse Spark 1.2 yet. Take the benchmarks as directional, not definitive.\n\n## The Gaps That Will Cost You Time\n\nWindows support requires WSL2. Given that roughly 42% of developers run Windows, this is not a footnote — it is a real barrier. The login flow requires a Facebook or Instagram account, which has generated backlash from developers who want no social media dependency in their toolchain. The contributor tier is also unavailable outside the US even through a VPN, cutting off developers in Canada, the UK, Australia, and most of Asia.\n\nBeta-quality issues are real: Docker sign-in failures, startup crashes, and inflated billing displays have all been reported. One documented case shows the agent researching a nonexistent project for three minutes before fabricating a complete plan. Long-horizon autonomy is where Muse Code is weakest right now.\n\n## Who Should Try It Now\n\nMuse Code earns a serious look for batch and high-volume workloads — PR triage, mechanical refactors, test generation at scale — where Standard tier pricing and parallel worktree agents deliver real cost and velocity gains. If you work with large repositories and do not need Windows support, the cost reduction over frontier alternatives is worth testing.\n\nWait if you need IDE integration, predictable monthly billing, or a mature plugin ecosystem. Do not put proprietary code through the contributor tier under any circumstances. And treat the beta label literally: this tool is nine days old. More [detailed setup guidance is available here](https://codersera.com/blog/muse-code-complete-guide-2026/) if you want to go deeper on configuration.", "url": "https://wpnews.pro/news/meta-muse-code-beta-what-developers-actually-get", "canonical_source": "https://byteiota.com/meta-muse-code-beta-what-developers-actually-get/", "published_at": "2026-08-14 09:08:15+00:00", "updated_at": "2026-08-14 09:34:58.367663+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-tools", "ai-products", "developer-tools"], "entities": ["Meta", "Muse Code", "Muse Spark 1.2", "Claude Opus 5", "GPT-5.6 Terra", "Grok 4.5", "Kimi K3"], "alternates": {"html": "https://wpnews.pro/news/meta-muse-code-beta-what-developers-actually-get", "markdown": "https://wpnews.pro/news/meta-muse-code-beta-what-developers-actually-get.md", "text": "https://wpnews.pro/news/meta-muse-code-beta-what-developers-actually-get.txt", "jsonld": "https://wpnews.pro/news/meta-muse-code-beta-what-developers-actually-get.jsonld"}}