{"slug": "meta-muse-code-check-your-tier-before-its-too-late", "title": "Meta Muse Code: Check Your Tier Before It’s Too Late", "summary": "Meta released Muse Code on August 5, a terminal coding agent running on Muse Spark 1.2, with a contributor tier priced at $0.10 per million input tokens and $0.20 per million output tokens, up to 21x cheaper than the standard tier, but the cheaper tier is the default and routes code into Meta's training pipeline. Developers who installed Muse Code in the last three weeks and skipped billing settings are already opted in, and Meta has not published a data retention window, processing regions, a DPA, SSO support, or an admin console. The standard tier costs $1.25 input and $4.25 output, and Meta's own benchmarks show Muse Spark 1.2 loses to Claude Opus 5 on all three, though comparisons were against GPT-5.6 Terra rather than the stronger Sol variant.", "body_md": "Meta shipped Muse Code on August 5 — a terminal coding agent running on Muse Spark 1.2, with a pricing structure that looks incredible at first glance: up to 21x cheaper than the standard tier. Here is what Meta does not lead with: the cheaper tier is the default, and it routes your code into Meta’s training pipeline. If you installed Muse Code in the last three weeks and skipped the billing settings, your proprietary code is already opted in.\n\nRun this before anything else:\n\n```\nmuse config set tier standard\n```\n\nNow let’s talk about what Muse Code actually is, what it gets right, and why most professional developers should stay on standard tier despite the price gap.\n\n## The Contributor Tier: 21x Cheaper, With a Catch\n\nMuse Code ships with two pricing tiers. The contributor tier costs $0.10 per million input tokens and $0.20 per million output tokens. The standard tier costs $1.25 input and $4.25 output — roughly 12–21x more depending on token mix. Meta frames this as developer choice. It is not, really, because the cheaper option is what you get by default.\n\nOn the contributor tier, you grant Meta permission to train future models on your prompts and completions. That means your code, your variable names, your architecture decisions, your internal API structures — all of it becomes training data. Meta has not published a data retention window, processing regions, a DPA, SSO support, or an admin console. Zero-data retention is listed as available “upon request through Meta sales,” which is enterprise negotiation language, not a self-serve toggle.\n\nThe risk is permanent. Once code is absorbed into model weights, no deletion request reverses it. Security teams at larger organizations have flagged this: default data collection settings move faster than governance policies, and developers evaluating a new tool rarely read the billing page before running it against a real project.\n\nTo opt out: verify your tier setting with `muse config tier`\n\n, then switch with `muse config set tier standard`\n\n. Confirm the change before pointing Muse Code at anything sensitive.\n\n## What Muse Code Gets Right\n\nThe architecture is worth understanding, because some of what Meta built here is genuinely novel.\n\nMost coding agents spawn helper subagents per task, then discard them. Muse Code keeps a set of background agents alive for the entire session — persistent agents for memory recall, skill recall, and goal tracking. This means the agent accumulates repository context instead of rediscovering it on every request. For long sessions across large codebases, this matters.\n\nWhen a complex task arrives, Muse Code spawns parallel sub-agents into isolated git worktrees under `.muse/worktrees/`\n\n. Each sub-agent gets a bounded scope and works independently, so parallel edits do not stomp on each other. Zuckerberg demonstrated this with six simultaneous feature builds — no collisions. Concurrency scales from 2 to 16 subagents depending on machine cores.\n\nEvery action enters an append-only JSONL event log stored locally. This makes crash recovery deterministic: `muse resume`\n\nreplays from the exact failure point — a real reliability improvement over agents that restart from scratch. The 1M-token context window holds dependency graphs, legacy code, and thousands of files in one session. In [Meta’s official launch post](https://developer.meta.com/ai/resources/blog/build-with-muse-code/), the team demonstrated Muse Code optimizing GPU kernels over 1,000+ tool calls across a 24-hour session on NVIDIA Hopper hardware — a compelling case for long-horizon engineering work.\n\n## The Benchmarks: Read the Fine Print\n\nMeta published its own benchmark comparisons, and to their credit, Muse Spark 1.2 loses on all three. That takes some honesty. The catch: Meta compared against GPT-5.6 Terra rather than Sol, the stronger variant. All numbers are vendor-run — no independent benchmarks exist yet for Muse Code.\n\n| Benchmark | Muse Spark 1.2 | Claude Opus 5 |\n|---|---|---|\n| Terminal-Bench 2.1 | 82.9% | 86.7% |\n| DeepSWE v1.1 | 59.3% | 65.0% |\n| Meta Internal Coding Bench | 70.6% | 79.4% |\n\nOne developer team extracted Muse Code’s system prompt and applied it to a rival model, reporting 2.7x fewer tokens and 2.4x lower cost for comparable results. That suggests at least some of the performance story lives in the prompt engineering. [Codersera’s independent analysis](https://codersera.com/blog/muse-code-complete-guide-2026/) also found discrepancies in earlier Muse Spark 1.1 benchmark claims, raising questions about reproducibility.\n\n## What’s Missing\n\nMuse Code is three weeks old, and it shows. No Windows support. No IDE extensions for VS Code or JetBrains. No subscription option — pay-per-token only. Sign-up requires a Facebook or Instagram account, which turned away a measurable portion of developers in the first week. Enterprise requirements — DPA, SSO, audit logs, admin console — are absent or on indefinite roadmap.\n\nMuse Spark 1.2 is also closed-weight. Meta built its developer reputation on Llama being free to download and self-host. Muse Spark goes the other direction: cloud API only, per-token pricing, with a data-for-discount tier as the default path. As [Verdent’s comparison guide notes](https://www.verdent.ai/guides/agents/what-is-muse-code), Claude Code and Codex CLI have roughly two years of production refinement; Muse Code has three weeks.\n\n## Who Should Use It Now\n\nThe contributor tier makes sense for one category: open-source projects where the code is already public. If you are building something MIT-licensed that lives on GitHub, the 21x price reduction is real money, and the data trade-off costs you nothing you were not already giving away.\n\nFor proprietary code, switch to standard tier and treat Muse Code as what it is — a beta from a first-time entrant in a crowded field. For enterprise compliance, wait. Meta has not shipped data processing agreements, audit logs, or zero-data retention guarantees, and is only beginning to take enterprise inquiries. [DevX’s coverage of the default data-sending behavior](https://www.devx.com/daily-news/muse-code-default-sends-data-to-meta/) puts it plainly: confirm your tier setting before pointing it at proprietary code.\n\nWorth watching. Not worth trusting with your code by default — and that default is the problem.", "url": "https://wpnews.pro/news/meta-muse-code-check-your-tier-before-its-too-late", "canonical_source": "https://byteiota.com/meta-muse-code-contributor-tier-privacy/", "published_at": "2026-08-25 11:08:51+00:00", "updated_at": "2026-08-25 11:16:23.262144+00:00", "lang": "en", "topics": ["ai-tools", "ai-products", "ai-policy", "ai-ethics", "developer-tools"], "entities": ["Meta", "Muse Code", "Muse Spark 1.2", "GPT-5.6 Terra", "Claude Opus 5", "NVIDIA Hopper"], "alternates": {"html": "https://wpnews.pro/news/meta-muse-code-check-your-tier-before-its-too-late", "markdown": "https://wpnews.pro/news/meta-muse-code-check-your-tier-before-its-too-late.md", "text": "https://wpnews.pro/news/meta-muse-code-check-your-tier-before-its-too-late.txt", "jsonld": "https://wpnews.pro/news/meta-muse-code-check-your-tier-before-its-too-late.jsonld"}}