Quick summary:• Meta released Muse Code in beta (macOS/Linux), a terminal coding agent powered by the Muse Spark 1.2 model. • Features repository-wide code generation, persistent background agents, and crash recovery via an append-only event log. • Standard Tier ($1.25/$4.25 per M tokens): Keeps your data private. • Contributor Tier ($0.10/$0.20 per M tokens): Up to 21x cheaper, but Meta trains future models on your prompts and code. • Meta’s reported benchmark numbers outpace independent tests, repeating a pattern seen with both Muse Spark Muse Spark 1.1 and 1.2 • Key evaluation factors for teams: Privacy trade-offs versus real-world performance compared to benchmark claims. |
|---|
What Is Muse Code
Muse Code is Meta's first terminal-based coding agent, released in beta on August 5, 2026. It installs with a single command on macOS or Linux and takes on software engineering tasks across entire repositories — planning a change, writing the code, and validating the result — rather than working snippet by snippet.
It's powered by Muse Spark 1.2, a coding-focused update to Meta's Muse Spark model family, which Meta co-trained specifically alongside Muse Code so the model and the harness fit together tightly.
How It Works: Background Agents and an Event Log
Two architectural choices define how Muse Code operates day to day.
Persistent async background agents. Rather than spawning a fresh subagent for each individual task, Muse Code keeps specialized background agents active for the whole session. They carry out next steps on their own and decide when to report back to the main agent — which cuts down on redundant information-gathering and reduces how often you need to steer it through a multi-step task.
An append-only event log. Every model call, tool run, approval, and edit gets appended to a local event log that acts as a single source of truth. Meta calls this "replay-exact and restart-safe" — if the agent crashes mid-task, it resumes exactly where it left off instead of losing progress or re-doing work.
Muse Code also ships with three default skills:
Skill | What It Does |
|---|---|
| Turns a task into an approval-gated plan before execution |
| Stress-tests that plan until it holds up |
| Works toward completing a specified objective end to end |
Muse Spark 1.2: The Model Behind It
Muse Spark 1.2 is a coding-focused update to Muse Spark 1.1, with Meta reporting gains in code generation, complex debugging, codebase understanding, and end-to-end developer workflows. Two training details stand out:
Co-training with the harness. Meta trained Muse Spark 1.2 using rejection-sampled trajectories from Muse Code itself, tuning the model to perform best specifically inside this tool — a move that mirrors an industry-wide shift toward treating models and their coding harnesses as a single product rather than separable pieces.
A self-improvement loop. Muse Spark 1.1 generated challenging coding environments and instruction-following templates, then graded candidate solutions against those requirements, producing training data for 1.2. Meta credits this loop with 1.2's improved ability to follow complex, multi-step instructions.
Benchmarks: What Meta Claims vs. What's Been Verified
This is worth reading carefully before you make any decisions based on Meta's launch charts.
On its own evaluation harness, Meta reports Muse Spark 1.2 scoring 82.9% on Terminal-Bench 2.1 and 59.3% on DeepSWE 1.1, positioning it in second place behind Claude Opus 5 on both benchmarks in its own comparison table.
Independent verification tells a more complicated story — and it's a pattern that's shown up before, not a one-off:
On the predecessor model, Muse Spark 1.1, Meta reported a Terminal-Bench 2.1 score of 80.0. Vals AI's independent testing measured 69.29 — a gap of more than 10 points. Separately, the Terminal-Bench team's own official verified leaderboard entry for Muse Spark 1.1 came in at 76.2%, still 3.8 points below Meta's claimed number, using the same harness family Meta says it used internally.
On Muse Spark 1.2, the raw score gap is narrower — independent evaluation from Artificial Analysis puts it around 80% against Meta's claimed 82.9%, roughly a 3-point difference. But the ranking gap is more striking: Meta's chart places Muse Spark 1.2 in 2nd place, just behind Claude Opus 5. Vals AI's common-harness leaderboard, by contrast, ranks it 14th.
For context, the highest independently verified Terminal-Bench 2.1 score posted by any model so far belongs to Claude Fable 5 running in Claude Code, at 83.8% — above Meta's own claimed (not independently verified) number for Muse Spark 1.2.
None of this means Muse Code doesn't work — beta users report it handles routine tasks like boilerplate, database queries, and API integrations well. But it does mean you shouldn't take Meta's launch-day comparison charts at face value, and it's worth checking independent leaderboards before routing production workloads its way.
Pricing: The Real Story
Muse Code's pricing is genuinely unusual, and it's arguably the most important thing to understand before adopting it.
Tier | Input (per 1M tokens) | Cached Input | Output (per 1M tokens) | Data Trade-off |
|---|---|---|---|---|
Standard | $1.25 | $0.15 | $4.25 | Meta does not train on your prompts or completions |
Contributor | $0.10 | $0.002 | $0.20 | Meta trains its models on your prompts and completions; capped at 60 requests per minute |
The Contributor tier is roughly 12.5x cheaper on input and 21x cheaper on output than Standard — for the identical model. That's not a loyalty discount; it's Meta pricing your code as training data. You select it by model ID rather than by signing any separate agreement, so the decision needs to be made deliberately, not defaulted into.
When Contributor tier makes sense: solo developers on personal projects, open-source work under permissive licenses, and non-sensitive prototypes where you're comfortable with Meta using the code to improve future models.
When it doesn't: any client work, proprietary codebases, code under NDA, or anything containing secrets, credentials, or PII. For agencies and regulated industries, this isn't really a pricing decision at all — it's a contractual one, and the answer is Standard tier by default until a client or compliance team says otherwise in writing.
Worth noting separately: Standard tier on its own is already competitively priced against flagship coding models from Anthropic and OpenAI, so the discount isn't the only reason to look at Muse Code — it's just the loudest one.
How It Positions Against the Field
Muse Code enters a category Anthropic (Claude Code) and OpenAI (Codex) have been building out for over a year, alongside independent players like Cursor. Meta's own launch materials position Muse Spark 1.2 just behind Claude Opus 5 on its internal benchmarks — a claim the independent verification above complicates. Meta itself acknowledges its evaluation harness may not be fairly tuned for third-party models, which cuts both ways: it's a reason to be cautious about Meta's own competitor comparisons in either direction.
A full, benchmarked comparison against Claude Code, Codex, and Cursor deserves its own dedicated piece once more independent evaluation data is available — this is a developing space and the rankings are still moving month to month.
Who Should Actually Use This
Worth testing now:
Solo developers and OSS maintainers who want cheap AI-assisted coding and are fine with the Contributor tier's data trade-off
Teams curious about the architecture (persistent background agents, crash-safe event log) who want to experiment on non-sensitive code
Worth waiting on:
Teams needing proven reliability on proprietary codebases today, given the gap between claimed and independently verified performance
Windows users — Muse Code supports macOS and Linux only at launch, with no confirmed Windows timeline
Anyone who needs training-data guarantees baked into a formal agreement rather than a tier selection
Limitations and Open Questions
Beta status — expect rough edges and rapid iteration
macOS and Linux only; no Windows support confirmed at launch
No independently verified benchmark leaderboard entry yet for Muse Spark 1.2 specifically, and the pattern from 1.1 suggests some caution is warranted
Not open source — weights aren't released, and access runs entirely through Meta's hosted API and Muse Code itself
The Contributor tier's 60-requests-per-minute cap will bottleneck any team trying to use it beyond individual, low-volume work
AIWerse Verdict
Muse Code's architecture — persistent background agents plus a crash-safe, replay-exact event log — is a genuinely useful design, and Meta co-training the model specifically for this harness is the right move for usability. But the two things that matter most for a real adoption decision both need a second look before you trust the launch headlines: the "up to 21x cheaper" Contributor tier is a data-training trade, not a free discount, and Meta's benchmark claims have consistently run ahead of independent verification across both Muse Spark 1.1 and 1.2. Worth testing on the Standard tier for non-critical work today. Worth waiting on independent benchmarks before moving proprietary, production workloads over.
FAQs
Is Muse Code free?
No. It has two paid tiers: Standard at $1.25/$4.25 per million input/output tokens with no data training, and Contributor at $0.10/$0.20 per million tokens where Meta trains on your prompts and completions.
What is the Muse Code Contributor tier?
A discounted pricing tier — roughly 12.5x cheaper on input and 21x cheaper on output than Standard — offered in exchange for letting Meta use your prompts and completions to train future models. It's capped at 60 requests per minute.
How does Muse Code compare to Claude Code?
Meta's own benchmarks place Muse Spark 1.2 close behind Claude Opus 5, but independent verification (including Vals AI's common-harness leaderboard) ranks it lower than Meta's launch materials suggest. A full benchmarked comparison needs more independent data before a confident call can be made.
Is Muse Code safe for private or client code?
Use the Standard tier, which Meta says does not train on your data. The Contributor tier should be reserved for open-source, personal, or explicitly non-sensitive work only — not client code, NDAs, or regulated data.
Does Muse Code run on Windows?
Not at launch. It currently supports macOS and Linux only, installed via a single terminal command.
Related News & Updates
Junaid Nawaz is the founder of AIwerse and a developer focused on AI tools, agentic workflows, and builder-focused tech. He covers AI model releases, coding tools, and platform updates for developers and teams building with AI. You can follow AIwerse on X (@AIwerse).