Meta's Muse Code Is Cheap Because You're the Dataset

Meta finally has a coding agent. Muse Code, a terminal harness in the mold of Claude Code and Codex CLI, hit beta on August 5 alongside Muse Spark 1.2, a coding-focused model update co-trained with the harness itself. The capability claims are plausible but thinly documented. The pricing, though, tells you exactly what Meta is really buying: your agentic trajectories.

A competent harness with one genuinely good idea

Architecturally, Muse Code is a fast follow. It's a single-command install (curl -fsSL https://dev.meta.ai/install.sh | bash, macOS and Linux only), it plans changes, edits code, and runs verification across large repos, and it ships three built-in skills — /plan for approval-gated planning, /grill for stress-testing a plan before execution, and /goal for driving toward a stated objective. If you've used Claude Code or Codex, nothing here will disorient you. That's the point: terminal agents have converged on a shape, and Meta built to spec.

Two design choices stand out. First, Muse Code runs async background agents that stay alive for the whole session instead of being spawned per task — Zuckerberg pitched it as a "multi-agent coding harness" that distributes work to sub-agents rather than making one agent do everything. That's a real departure from the spawn-on-demand subagent pattern most harnesses use, and it should cut the latency tax you pay every time a fresh subagent re-reads context.

Second, and better: every model call, tool run, approval, and edit is appended to a local event log, which Meta says makes the runtime "replay-exact and restart-safe." Crash mid-refactor and the agent resumes exactly where it stopped. This is durable-execution thinking — the idea behind workflow engines like Temporal — applied to coding agents, and it addresses a failure mode every heavy agent user knows: the two-hour session that dies at 90% and can't be reconstructed. Expect competitors to copy this within a quarter.

The model side follows the industry's co-training playbook. Muse Spark 1.2 was trained on long-horizon coding work — whole-repository generation, end-to-end projects — including rejection-sampled trajectories from the harness itself, plus a reported 1 million-token context window. Anthropic and OpenAI already tune their models against their own harnesses; Meta is just saying the quiet part in the release notes. The harness is no longer a wrapper around the model. It's part of the model.

What Meta didn't publish is a clean eval story. The announcement leans on charts for Terminal-Bench 2.1 (89 tasks), DeepSWE v1.1 (113 tasks across 91 repos), and a 440-task internal bench, without a methodology write-up, and Hacker News commenters were quick to note the comparison charts pit Spark 1.2 against a rival's mid-tier model — and still drop some rows. One HN commenter claimed that, until recently, "everyone at Meta was using Claude Code" internally. Treat the capability claims as unproven until third-party Terminal-Bench numbers land.

The contributor tier is the actual announcement

Standard Muse Spark 1.2 API pricing is unremarkable: $1.25 per million input tokens, $0.15 cached, $4.25 output — unchanged from Spark 1.1, parked deliberately in Sonnet/GPT-mid territory.

The contributor tier is the strategy. Opt in to letting Meta train on your sessions and the price collapses to $0.10 per million input tokens, $0.002 cached, and $0.20 output — a 12x discount on input and 21x on output. At those rates a heavy day of agentic coding costs pocket change, undercutting every hosted competitor and landing in DeepSeek's neighborhood with (Meta claims) frontier-adjacent capability.

Look at what Meta gets in return. Agentic coding trajectories — full sequences of plan, tool call, error, correction, and verified fix inside real repositories — are the scarcest, highest-value training data in the industry right now. Every lab is trying to manufacture them synthetically or harvest them from consenting users. Google seeded Gemini CLI adoption with a free consumer tier that can feed model improvement; Meta just put an explicit sticker price on the same trade. This is the classic Meta move, translated to 2026: the social products were free because the graph was the product. Muse Code is nearly free because your debugging session is the product. It isn't a discount — it's a data acquisition, and an efficiently priced one.

The open-weights company didn't show up

Here's the part that should sting if you bet your stack on Meta's old identity: Muse Spark 1.2 has no downloadable weights. Hosted API only. The company that spent three years arguing open weights would win the ecosystem — and whose Llama releases anchored the entire self-hosting movement — has shipped its flagship coding model as a closed, hosted dependency, completing the retreat that began when Superintelligence Labs formed and Zuckerberg started hedging on what Meta would keep open-sourcing.

Muse Code's own state is local — the event log lives on your machine — but calling this a local-first toolkit stretches the term past breaking. If Llama's openness was why Meta had your goodwill, that goodwill is now unsecured debt.

Should you switch?

Depends entirely on which trade you're willing to make.

Solo developers, students, side projects: the contributor tier is close to irresistible if you can stomach the terms. Nothing hosted comes close per token, trying it costs one curl command, and terminal agents are cheap to trial precisely because they've converged — your muscle memory transfers. Just be honest with yourself that every repo you point it at becomes training data.

Enterprises should wait. It's a beta, sign-in runs through Meta's developer platform (HN commenters flagged both corporate firewalls and the general allergy to tying work tooling to Meta accounts), and zero-data-retention is at the "starting to accept requests" stage rather than a checkbox. Anthropic and OpenAI are simply further along on the boring compliance machinery that enterprise adoption actually rides on.

The sharpest way to read this launch: Muse Code doesn't need to beat Claude Code today. It needs to harvest enough trajectories at $0.20 per million output tokens to make Muse Spark 1.3 the model that does. The capability race is unresolved; the data play is already live. The question worth watching isn't whether Meta's benchmarks hold up — it's whether Anthropic and OpenAI decide they can't afford to leave the pay-with-your-data tier unanswered.