Meta released Muse Spark 1.2 and the agent Muse Code at 20 cents per million tokens, acknowledges shortcomings in its own benchmarks

Meta released the coding-focused model Muse Spark 1.2 along with its first in-house terminal coding agent, Muse Code, currently in beta. According to the company, the model invests more compute in training on programming tasks, plans steps ahead, tracks a fixed goal and compresses its previous context in long sessions instead of truncating it. Muse Code can resume work after a crash at the exact point of interruption instead of reloading the entire context, because it logs every model call, approval and change to a local log file. Meta also acknowledges methodological shortcomings in its comparative tests against competing models—the test environment was not optimized for competitors, the result achieved elsewhere by the model Claude Opus 5 is two percentage points higher according to other leaderboards, and the model Kimi K3, which the company tested according to its own methodology, is entirely absent from the published results.

According to Meta, Muse Spark 1.2 primarily improves coding capabilities over the model Muse Spark 1.1 released earlier this year, with advances in code generation, debugging and reasoning about large codebases. The model was trained mainly on long-running tasks, such as generating entire repositories or conducting independent research, and some of the training data was created by its predecessor Muse Spark 1.1 itself, which generated programming tasks and evaluated the quality of solutions.

According to Mark Zuckerberg, CEO of Meta, Muse Code can handle end-to-end software engineering across large repositories—planning changes, writing code and verifying results. It installs with a single command, similarly to Claude Code or Codex, and offers the commands “/plan” for plan approval, “/grill” for checking the plan for weaknesses and “/goal” for tracking a fixed goal. For larger tasks, the agent launches concurrent sub-agents working in isolated working copies without collisions; according to Zuckerberg, it built six game features in parallel without conflicts in a test using this approach.

The standard pricing plan remains at the level of version 1.1 (1.25 dollars per million input tokens and 4.25 dollars per million output tokens), with a new plan priced at 20 cents per million output tokens, conditional on sharing user data for further model training. According to the sources, Western competitors charge 10 to 30 dollars per million output tokens, Chinese providers start at 18 cents, and the model Kimi K3 costs 3 dollars per million input tokens and 15 dollars per million output tokens (on a cache hit, the input price drops to 30 cents). Alexandr Wang, head of the AI division at Meta, said the company competes primarily on price, not capabilities. The sources also state that shares in Meta fell by ten percent last week and that the company generates 98 percent of its revenue from advertising.