Muse Code — Meta's $0.10 Bet That You'll Trade Code for Tokens

Meta banned Claude Code and Codex internally on June 29 over distillation risk — then shipped Muse Code on August 5, a terminal coding agent that asks your engineers to hand their code to Meta’s training pipeline in exchange for a discount. That is not hypocrisy. It is a strategy with a price tag attached: $0.10 per million input tokens on the contributor tier versus $1.25 on standard. The discount is the product. Your code is the payment.

TL;DR

• What: Meta launched Muse Code in beta — a terminal coding agent powered by the closed-weights Muse Spark 1.2 model
• Pricing play: Contributor tier at $0.10/$0.20 per million tokens (input 12x cheaper, output 21x cheaper than standard) — in exchange for letting Meta train on your code
• Closed-weights pivot: Third Muse Spark release in four months, none with open weights or Hugging Face downloads — Llama-era openness is dead
• Missing: No SWE-bench, tbench.ai, or any independent benchmark scores at launch

What Happened

Muse Code is Meta Superintelligence Labs’ first coding agent, released August 5 in beta for macOS and Linux. It runs in the terminal, powered by Muse Spark 1.2 — the third model in a sequence that started with Muse Spark in April, followed by 1.1 in July, and now 1.2 in August. All three are closed-weights. No Hugging Face repository. No parameter count. No architecture disclosure. No self-hosting.

The pricing structure splits into two tiers. Standard runs $1.25 per million input tokens and $4.25 per million output tokens, which places it roughly in line with competing frontier model APIs. The contributor tier drops to $0.10 input and $0.20 output — making input about 12.5x cheaper and output more than 20x cheaper. The difference is not a volume discount. It is a data licensing agreement: contributors allow Meta to use their interactions, including code, for model training.

The contributor tier is rate-limited to 60 requests per minute and 2.1M tokens per minute. Standard gets 3,000 requests per minute and 4M tokens per minute. For anything beyond hobby use, the throughput ceiling matters as much as the per-token price.

The agent architecture has two features worth noting. First, Muse Code runs persistent async background subagents that stay alive for the entire session rather than spawning and dying per task. Workers can investigate a repository, prepare next steps, and return findings without cold-starting. Second, every model call, tool run, approval, and edit is appended to a local event log that Meta calls restart-safe — if a task crashes, the agent resumes from the last checkpoint rather than replaying from scratch.

Meta’s marquee case study claims Muse Spark 1.2 iteratively optimized GPU kernels over more than 1,000 tool calls in runs lasting up to 24 hours. They also report six simultaneous game features implemented with no collisions. Both are internal results with no independent reproduction.

Why This Matters

The contributor tier is a data acquisition funnel disguised as a pricing page. Meta needs proprietary coding data to close the gap on Anthropic and OpenAI in code generation quality, and they are willing to subsidize access to get it. The math is transparent: at $0.10/$0.20, Meta is almost certainly losing money on inference. The value extraction happens downstream, when your code improves Muse Spark 1.3 or whatever comes next.

This matters because of timing. Meta explicitly banned its own engineers from using Claude Code and Codex to prevent competitors from training on Meta’s internal code. The policy was documented around June 29 — five weeks before Muse Code launched with a tier that does exactly what Meta feared from the other direction. Meta understands the value of proprietary training data because they are actively protecting their own. The contributor tier is Meta asking you not to extend the same courtesy to yourself.

The closed-weights pivot is the second story here. Muse Spark 1.2 is the third release from Meta Superintelligence Labs in four months, and none have followed the Llama playbook. No open weights. No community fine-tuning. No self-hosting path. Meta has forked its AI strategy: Llama for the ecosystem play, Muse for the competitive play. If you built workflows around the assumption that Meta would keep releasing open-weights frontier models, Muse is the correction. The models Meta believes give it a competitive edge stay locked down.

The co-training architecture is genuinely interesting from a technical standpoint. Muse Spark 1.2 was trained inside its own agent runtime, with Muse Spark 1.1 generating training environments and grading candidate solutions. The harness and model were optimized as one unit using rejection-sampled trajectories. This is a different approach from bolting a model onto an existing agent framework — the model and the runtime evolved together. Whether that translates into real-world task completion superiority over Claude Code or Codex is an open question, because Meta has published zero independent benchmark results.

Compare the benchmark gap directly: Claude Code on Opus 5 and OpenAI’s Codex have published SWE-bench scores. Muse Code has internal case studies and nothing else. Until someone runs SWE-bench Verified or tbench.ai on Muse Spark 1.2 independently, you are evaluating marketing materials, not engineering evidence.

The persistent background agents and crash-safe logging are solid architectural choices that address real pain points in long-running coding sessions. Anyone who has watched a 200-step Claude Code session fail at step 180 and replay from scratch knows the value of checkpoint-based recovery. But these are infrastructure features, not model quality features. They make the tool more reliable without telling you whether the model behind it produces better code.

The Take

I would not touch the contributor tier in any professional context. The 12-21x discount is real, but you are paying with something more expensive than tokens: your proprietary code enters Meta’s training pipeline, where it improves a closed-weights model you cannot inspect, cannot self-host, and cannot fine-tune. You are subsidizing your own future competitor. For hobby projects and open-source work where the code is already public, the contributor tier is defensible — you are not giving up anything that is not already available. For anything with proprietary value, the standard tier at $1.25/$4.25 is the only rational choice.

The standard tier itself is a harder call. The pricing is competitive, and the persistent subagents plus crash-safe logging solve real workflow problems. But Muse Spark 1.2 is a model with no published architecture, no parameter count, and no independent benchmark validation. You are betting entirely on Meta’s internal evaluations. Claude Code and Codex have public scores you can compare against independent reproductions. Muse Code asks you to trust a case study about GPU kernel optimization and take it on faith.

My recommendation: do not migrate any production workflows to Muse Code until independent SWE-bench Verified or tbench.ai results exist. If you want to evaluate it, use the standard tier on non-proprietary codebases and compare output quality against your current agent on identical tasks. The architecture is promising. The pricing strategy is transparent. The model quality is unproven. Act accordingly.