Meta's Muse Code: a terminal coding agent worth a slot in your eval bake-off

What it is

Meta's first coding agent, shipped in beta on August 5, 2026. Muse Code is a terminal-based AI coding agent designed to help developers manage complex tasks across large software repositories, powered by the new Muse Spark 1.2 model. It's a direct entrant into the Claude Code / Codex / Copilot CLI category — not an IDE, not a chat window, a CLI that plans, edits, and reviews.

The next-work-session test

Concrete scenario: you have a 400-file service, a flaky integration test, and a refactor ticket you keep deferring. Muse Code's pitch is that you can hand it the ticket, let it plan, then let it work in the background while you keep coding. Meta says model calls, tool runs, and code edits are recorded in a local event log, allowing interrupted sessions to be resumed with the muse resume command. The tool also uses isolated Git worktrees to support parallel tasks. For anyone already running an agent loop with subagents and worktrees, that's a familiar shape — the interesting part is whether Muse Spark 1.2 holds up on your repo, not the vendor's. Slot it into your existing eval harness alongside Claude and see.

Pricing

Verified via Meta's own launch coverage and secondary reporting. Muse Code launched in beta as a terminal AI coding agent. Standard API prices are $1.25 input, $0.15 cached and $4.25 output per million tokens. There's also a data-for-discount tier: a lower-cost contributor tier charges $0.10 per million input tokens, $0.002 per million cached input tokens, and $0.20 per million output tokens in exchange for allowing Meta to use contributor data to improve its models. Meta's AI chief Alexandr Wang characterizes the contributor tier as "more than 10 times cheaper than than even the pay-as-you-go tier." That's Meta's framing, not an independent measurement — the actual savings depend on your input/output mix.

What we'd actually use it for

Honestly? A fourth column in your agent bake-off spreadsheet. If you already run evals across Claude Code, Codex, and Copilot, adding Muse Code costs you a script and an API key. The two things worth testing specifically:

The contributor tier is genuinely interesting for hobby projects and OSS work where your code is already public. It's not interesting for anything under NDA.

Limits

The benchmark story is middle-of-the-pack, not leading. Muse Code runs in the terminal, finishes second to Claude on benchmarks, and offers a contributor tier that is 21 times cheaper than standard pricing. The catch: Meta gets to train on your code. On Meta's own reported numbers, Muse Spark 1.2 scored 77.4 on SWE-Bench Verified against Claude Opus 4.6's 80.8 and Gemini 3.1 Pro's 80.6, and lagged well behind GPT-5.4 on GDPval. Close, not ahead.

Platform coverage is thin: no Windows support, no GUI, and no IDE plugin at launch. And the enterprise story is essentially absent — as of early August 2026 there is no published retention window, no stated processing region, no SSO or admin console, no server-side audit log, no DPA and no scoped compliance attestation for the Model API. If procurement needs to sign off, this isn't ready. Also worth flagging: benchmark comparisons here are model-in-its-own-harness, which is not apples-to-apples — treat the leaderboard as directional only.

Try it if

• You already maintain an eval harness and can drop a new agent into it in an afternoon.
• You work in a terminal-first loop with worktrees and background tasks.
• You have OSS or personal projects where the contributor tier's data terms are acceptable.
• You want a cheaper option to burn tokens on speculative refactors before spending Claude money.

Skip it if

• Your code is proprietary and you're tempted by the contributor tier — read the data-use terms first, and probably don't.
• You need Windows, an IDE plugin, or a GUI. None exist yet.
• You need SOC 2 scoping, SSO, DPAs, or audit logs for the Model API.
• You want the current benchmark leader for a critical migration — that's still Claude on most public numbers.

Install command and full spec are in Meta's launch post; pricing details and the contributor tier trade-off are broken down further in this pricing guide and TechCrunch's launch coverage.