Meta's First Coding Agent Is 95% Off, If Meta Can Train on Your Code

TL;DR

Meta entered the coding-agent wars on August 5 with Muse Code, a terminal agent in beta for macOS and Linux, powered by a co-trained Muse Spark 1.2 model. The architecture is legitimately novel: persistent async background agents, sub-agents fanned out into isolated git worktrees, and an event log that makes every session replay-exact. But the part that will actually move the market is the pricing. The standard API tier runs $1.25 per million input tokens and $4.25 output; a "contributor" tier drops that to $0.10 and $0.20, roughly 12x and 21x cheaper, in exchange for explicit permission to train future Meta models on your prompts and completions. Meta spent two decades perfecting free products where you are the product. This is the first time it has published the exchange rate.

Meta finally shows up to the terminal

Muse Code is Meta's first coding agent, and it lands squarely in territory owned by Claude Code and OpenAI's Codex. Per TechCrunch, it takes on complete software engineering tasks across large repos: planning changes, writing code, validating the results. Alexandr Wang, who runs Meta Superintelligence Labs, pitched it as "an incredibly good option, especially from a cost perspective." Subtle.

Install is a one-liner: curl -fsSL https://dev.meta.ai/install.sh | bash. It is beta, terminal-only, and per 9to5Mac there is no dedicated app interface yet, unlike its two rivals. Windows users are not invited.

The architecture is the interesting part

Most agent harnesses spawn a fresh helper per task and throw it away. Muse Code instead runs a simple main loop plus a set of persistent async background agents that stay alive for the whole session, carry out next steps on their own, and decide when to report back to the main agent.

Underneath sits a local event log: every model call, tool run, approval, and edit is appended to the log before it executes, which Meta says makes sessions replay-exact and restart-safe. It is the same trick databases have used for decades with write-ahead logs: commit the intention to the tape before you act, and a crash becomes "replay the tape" instead of "start over and hope." Your agent dying at hour three of a refactor stops being an incident and becomes a resume.

Three built-in skills round it out: /plan for approval-gated planning, /grill for stress-testing a plan before you trust it, and /goal for autonomous task completion. When a job is big enough, the agent fans out to sub-agents working in parallel, each in its own isolated git worktree. Meta's line: "Your working copy is never touched," and in testing it had the agent build six features for a game simultaneously with no collisions.

The pricing is the actual story

Muse Spark 1.2 comes through the Meta Model API in two tiers, and the spread between them is the whole announcement. Standard: $1.25 per million input tokens, $0.15 cached input, $4.25 output, with a commitment that your prompts and completions are not used to train Meta's models. Contributor: $0.10 input, $0.002 cached, $0.20 output. That is roughly 12x cheaper on input, 21x on output, and 75x on cache reads, per Unite.AI and BigGo.

The catch is right there in the tier name. Contributors explicitly permit Meta to train future models on their prompts and completions. For a coding agent, "prompts" is not your chat small talk; it is your repository, streamed to Meta as context, at up to a million tokens per request going by the Vercel AI Gateway listing. Reporting also notes the contributor tier carries much tighter rate limits than standard, which tells you who it is for: individuals and experiments, not production fleets.

Muse Spark 1.2: co-trained, benchmark-shy

The model itself is a coding-focused update with, in Meta's words, significantly scaled-up training compute on coding tasks and broader training-environment diversity: whole-repository generation, large end-to-end projects, long-horizon work. The notable design choice is that Muse Spark 1.2 was co-trained with Muse Code, so the model and harness were optimized as a unit rather than the usual generic-model-plus-wrapper arrangement. Meta also describes a self-improvement loop in which the previous Muse Spark generated challenging coding environments for the new model to train against.

What Meta did not do is publish serious numbers. Per Unite.AI, the launch post shows charts for Terminal-Bench, DeepSWE, an internal coding benchmark, and GDPval as images, with no methodology write-up, and no new SWE-Bench Verified score for 1.2. Until independent evals land, treat the capability claims as marketing and the architecture as the substance.

What this means for you

First, the economics. Claude Code and Codex subscriptions are priced against API rates an order of magnitude above Meta's contributor tier. If Muse Spark 1.2 is even in the same capability neighborhood, a $0.10/$0.20 agent loop changes the math for high-volume, low-stakes work: batch refactors, test generation, throwaway prototypes. Expect pricing pressure on everyone.

Second, the data terms deserve a straight-faced read. On standard tier, Meta commits to not training on your traffic. On contributor tier, opting in means your codebase context becomes training data. That is a fine trade for open-source and side projects, and a firing-offense trade for your employer's monorepo. Check your org's policy before someone pipes the install script into bash on a work laptop.

Third, the harness ideas are worth stealing regardless of whose model you run: an append-before-execute event log, long-lived background agents, and worktree-isolated parallelism are all patterns that make agents auditable and crash-tolerant, and none of them are Meta-specific. The model is also live on OpenRouter if you want to try it without touching the tooling.

Key Takeaways

• Meta launched Muse Code on August 5: its first coding agent, terminal-only, in beta for macOS and Linux, powered by a co-trained Muse Spark 1.2.
• Architecture highlights: persistent async background agents, an append-before-execute event log that makes sessions replay-exact and restart-safe, and parallel sub-agents in isolated git worktrees.
• Standard API pricing is $1.25/$4.25 per million tokens (input/output) with a no-training commitment; the contributor tier is $0.10/$0.20 in exchange for training rights on your prompts and completions.
• The contributor tier reportedly carries much tighter rate limits, aiming it at individuals and experiments rather than production.
• Meta published no methodology for its benchmark charts and no new SWE-Bench score; capability claims are unverified until independent evals land.
• Never point the contributor tier at proprietary code; the discount is paid for with your repository as training data.

Sources: Meta (official announcement), TechCrunch, 9to5Mac, Unite.AI, BigGo, Vercel AI Gateway