Meta ships Muse Code agent on new Muse Spark 1.2 model
Meta Superintelligence Labs released a terminal coding agent in beta alongside a co-trained model priced at $1.25 per million input tokens, undercutting rivals.
Meta Superintelligence Labs released Muse Code, a terminal-based coding agent, in beta on Wednesday, paired with Muse Spark 1.2 — a model the company says was co-trained with the agent harness rather than adapted to it afterwards.
Architecture
Muse Code installs via a single command on macOS and Linux. Its distinguishing design choice is persistent asynchronous background agents that stay alive for the duration of a session, instead of subagents spawned and discarded per task. Meta says this cuts repeated context-gathering and reduces the steering a developer must supply on long multi-step work. When a task is complex enough, the agent parallelises across isolated git worktrees, leaving the user's working directory untouched; internal testing built six game features concurrently without merge conflicts.
The runtime writes every model call and action to a local append-only event log, which allows replay-exact resumption after a crash. Built-in skills include /plan for approval-gated planning, /grill for adversarially stress-testing a plan before execution, and /goal for objective tracking.
Model and pricing
Meta scaled up training compute specifically on coding tasks for Muse Spark 1.2. Its own launch charts place the model second behind Claude Opus 5 on Terminal-Bench 2.1 (89 tasks), DeepSWE v1.1 (113 tasks across 91 repositories and five languages) and an internal benchmark of 440 tasks derived from real pull requests — with the caveat, which Meta states, that its harness may not be tuned for competitors' models. In a long-horizon test the agent optimised KDA and MLA GPU kernels on Nvidia Hopper hardware across more than 1,000 tool calls over roughly 24 hours.
API pricing is $1.25 per million input tokens and $4.25 per million output. Meta also introduced a contributor tier at $0.10 and $0.20 — a 12.5x and 21.25x discount — in exchange for permission to train future models on submitted prompts and completions, capped at 60 requests per minute versus 3,000 on the standard tier. Weights are not published; this is a hosted dependency.
Why it matters
Meta now competes directly with Claude Code and OpenAI's Codex, and it is competing on price. Independent evaluation put Muse Spark 1.2 at roughly $0.40 per Intelligence Index task, among the cheapest at its capability level. The data-for-discount tier is the more novel move: it converts developer telemetry into an explicit line item, which enterprise buyers will have to price against their own code-confidentiality requirements.