Every figure, date and product claim the video states, chased to a source.
Muse Code is a terminal coding agent from Meta, released in beta on 5 August 2026, powered by Muse Spark 1.2. Meta describes it as taking on “complex software engineering tasks across large repositories: planning changes, writing code, and validating the results” and as running “a simple agent loop plus a set of async background agents”. It installs on macOS or Linux.
It keeps a local event log. Every model call, tool run, approval and edit is appended to it, which Meta says makes the runtime “replay-exact and restart-safe”: after a crash the agent resumes exactly where it stopped.
It ships with three default skills: /plan, which turns a task into an approval-gated
plan; /grill, which stress-tests that plan; and /goal.
It installs and runs from a single command, and is positioned against OpenAI’s Codex and Anthropic’s Claude Code on cost.
Muse Spark 1.2 is a coding-focused update to Muse Spark 1.1, with improvements Meta lists in code generation, complex debugging, codebase understanding and end-to-end developer workflows. Meta says it scaled up training compute on coding tasks and trained the model on long-horizon work including whole-repository generation.
Context window: 1 million tokens.
Muse Spark 1.2, per million tokens:
| Tier | Input | Output |
|---|---|---|
| Standard | $1.25 | $4.25 |
| Contributor | $0.10 | $0.20 |
That is a 92% reduction on input and 95% on output, or roughly 12.5x cheaper for input and 21.25x cheaper for output.
The Contributor tier is the discount in exchange for training rights. Enabling it grants Meta permission to use the user’s prompts and completions to train future models.
The Ranking Engineer Agent (REA) autonomously generates hypotheses, launches training jobs, debugs failures and iterates on results across Meta’s ads-ranking work. REA-driven iterations doubled average model accuracy over baseline across six models, and three engineers delivered proposals for eight models, work that had historically taken two engineers per model.
Unified AI agents for capacity efficiency at hyperscale. The platform pairs MCP tools (standardised interfaces for querying profiling data, code and configs) with “skills” encoding expertise from senior engineers. Meta reports the systems have recovered hundreds of megawatts of power and compressed roughly ten hours of manual investigation into about thirty minutes.
Swarms of specialised agents mapping tribal knowledge. A swarm of 50+ agents read every file across 4,100+ files in three repositories and produced 59 context files encoding knowledge that had lived only in engineers’ heads. Structured navigation guides went from 5% of code modules to 100%, more than 50 non-obvious patterns were documented, and preliminary tests showed 40% fewer agent tool calls per task. The swarm is split by role: explorer, module analyst, writer, critic and fixer agents.
Claude Code (Anthropic), Codex (OpenAI), Gemini in Google’s developer tooling, Cursor and GitHub Copilot are all shipping coding tools at the time of writing, and Muse Code is positioned directly against Claude Code and Codex.