Meta trades discount for training data: Muse Spark 1.2 lures developers with cheap rate

Meta has released Muse Spark 1.2, primarily a coding update of Muse Spark 1.1. The company cites improvements in code generation, debugging, and understanding large codebases, achieved by investing more compute in programming tasks and increasing training environments. A portion of training data was generated by the predecessor model itself: Muse Spark 1.1 created tasks and instruction templates and evaluated how well solutions met them. Benchmarks mentioned include Terminal-Bench 2.1, DeepSWE v1.1, and 440 internal tasks, compared with Grok 4.5, Claude Opus 5, GPT-5.6 Terra, and Gemini 3.6 Flash, showing visible improvement but not always matching the top. Meta notes that the test setup is not tuned to other models. The methodology mentions Kimi K3, but it is missing from published benchmarks; on Terminal-Bench 2.1, K3 scores close to Opus 5 and above Spark 1.2. DeepSWE runs are not comparable because each model ran in its own agent. Meta also introduced Muse Code, a terminal-based agent similar to Claude Code and OpenAI Codex, with a planning mode and a unique '/grill' command to critique plans. It keeps subagents active for the whole session and can resume work after crashes with high accuracy. The standard rate is unchanged at $1.25 per million input tokens and $4.25 per million output tokens, but a new rate of 20 cents per million output tokens requires users to share data for training. Western competitors charge $10-30 for output, Chinese providers from 18 cents.