Muse Code and Muse Spark 1.2

That pricing move shaped the whole read on the launch. People saw it as transparent on its face, but also as a blunt statement about how valuable training data is. Others immediately distrusted it, arguing there is no practical way to verify what Meta does with data regardless of billing tier. The result is that Spark 1.2 landed less as “here is the new best coding model” and more as “here is a cheap enough option to force a fresh privacy versus cost tradeoff.” Several comments also noted that the cheap tier appears geographically limited, with the EU excluded. On model quality, the reaction was skeptical. Readers zeroed in on Meta’s benchmark framing, especially the choice to compare against OpenAI’s mid-tier Terra in prominent charts while omitting stronger top-tier comparisons from the main marketing. Some pointed out that the page does include Sol lower down, where Spark 1.2 loses, and that Opus is included and still wins most benchmarks. That left a broad impression that Meta is closer to the middle of the pack than to the frontier, and that the post is using selective presentation to soften that fact. Benchmarks themselves did not get dismissed outright. The more grounded view was that they remain useful as one comparable signal, but vendor-picked benchmark sets are now marketing material first. Muse Code itself drew interest mostly as a harness, not as a reason to switch. A few comments noted practical features like running workers in separate worktrees and crash recovery. Another commenter poking at the binary said it appears to be written in Rust and looks heavily inspired by Codex CLI, though not obviously a direct fork. The underlying assumption across the conversation was telling: if Meta employees are still choosing Claude Code or Codex when they can, then Muse Code has not yet earned default-tool status on capability alone. The mood was cautious and cynical, but not dismissive. People do see a usable release here. They just do not see evidence that Meta has caught the best coding models. The compelling part is the economics. If you can tolerate retention, Spark 1.2 may be a real alternative for personal projects or low-sensitivity workloads. If you cannot, the launch mostly reinforces how much trust, data handling, and benchmark honesty now matter in model selection.