Note: This post was written by Muse Spark 1.2. The following is a synthesis of reporting from major news organizations and Meta’s research blog.
On August 5, Meta shipped two things it says were built as a pair: Muse Spark 1.2, a coding-focused update to its Muse family, and Muse Code, a terminal agent that runs on it. The combo is Meta’s first serious bid for paid developer workflows — the territory Anthropic’s Claude Code claimed in late 2025 and OpenAI’s Codex now contests.
A Coding Update, Not a New Line
Muse Spark is the series from Meta Superintelligence Labs, the group that rebuilt Meta’s model stack over nine months beginning in 2025. It debuted in April 2026, then Muse Spark 1.1 arrived July 9 as a preview for agent and coding use. Meta describes each step as deliberate scaling — small and fast first, validating before going larger.
Muse Spark 1.2 follows that pattern. It is not a new line but a coding-tuned step on 1.1: more training compute on coding tasks, broader environment diversity, and the same emphasis on first-attempt accuracy and reliable tool calling for long agentic sequences. The window remains 1 million tokens.
Meta highlighted two training choices that shape how the model behaves in the agent:
Co-training with the harness. Instead of training the model and then dropping it into Muse Code, Meta trained them together — rejection-sampled harness trajectories, recipe work for goals, compaction, and subagents, and toolset integration. The aim is that Spark 1.2 is best when paired with the agent it will actually run in.
A self-improvement loop. The team used Spark 1.1 to generate hard coding environments and instruction templates, then had it grade candidate solutions. Those graded pairs became scalable training data for 1.2, targeted at following complex, multi-step instructions.
Muse Code: Built for Long Runs
Muse Code is a command-line agent for macOS and Linux:
curl -fsSL https://dev.meta.ai/install.sh | bash
It takes on what the field now calls agentic coding: plan changes across a large repo, write code, and validate the result, with the human approving rather than steering each edit.
Three design decisions set it apart from a chat-style helper:
Persistent background agents. Rather than spawning a helper for one subtask and tearing it down, Muse Code keeps specialized async agents alive for the whole session. They share context, pick next steps themselves, and choose when to report back. Meta says this cuts repeated information gathering and reduces prompting on multi-step work.
Parallel subagents in isolated worktrees. Large jobs fan out to concurrent subagents, each in its own isolated worktree. That is one workspace per agent, so parallel edits do not collide — the same isolation that makes local multi-agent tests safe.
A local event log. Every model call, tool use, approval, and edit is appended to a local log. That log is the runtime’s source of truth: replay-exact and restart-safe. After a crash, the agent resumes where it stopped. It is a durability choice aimed at jobs that run for hours.
Three bundled skills sit atop that loop: /plan builds an approval-gated plan, /grill pressure-tests it, and /goal drives toward the stated objective until it is met.
The Thousand-Call Test
To demonstrate long-horizon behavior, Meta pointed to GPU kernel optimization. Running through Muse Code, Spark 1.2 wrote, compiled, profiled, and iteratively improved KDA and MLA kernels for NVIDIA Hopper across more than 1,000 tool calls over up to 24 hours.
The baseline was FLA’s Triton implementation of KDA. Importing optimized libraries was barred — the model had to implement in Triton and apply kernel knowledge directly. Meta said the agent paired a chunk-parallel preparation kernel with a sequential inter-chunk scan, using fusion and tiling plus KDA-specific moves like re-centering a gated cumulative decay at the chunk midpoint, and kept improving over the baseline throughout the run.
The kernel matters less than the pattern: write, measure, rewrite, sustained for a day without losing direction. Meta credits planning, goal conditioning, and context compaction for holding course that long.
Price and Access
Both model and agent are available now — Muse Code in beta, Spark 1.2 inside it, plus the Meta Model API and OpenRouter.
Pricing is the aggressive part, and the trade is explicit:
| Tier | Input / Output per million | Data retention | Rate limits |
|---|---|---|---|
| Contributor | $0.10 / $0.20 | Meta may retain prompts and completions to train future models | 60 req/min, 2.1M tokens/min |
| Standard | $1.25 / $4.25 | Not retained for training | 3,000 req/min, 4.0M tokens/min |
The standard numbers match Spark 1.1’s July pricing, which Meta’s leadership described as built to scale with heavy use. The contributor tier is less than one-tenth the output cost — $0.20 versus $4.25 — but requires opting in to feedback and training use.
Muse Code itself is pay-as-you-go on token use, no seat license in beta. For comparison, most coding agents this year cluster in the low single digits per million tokens; Meta’s $0.10 input sits at the floor.
Where It Stands
The framing in coverage has been consistent: Muse Code is Meta’s equivalent to Codex and Claude Code, not another chat model. Reuters and business press positioned it as Meta’s first major agentic coding product; developer press focused on the co-training and durability details.
The technical bet is to pair model and harness, persist state so long sessions survive failure, and parallelize with isolation. The product bet is price: buy experimentation on the contributor tier while keeping a standard tier for teams where retention is a non-starter.
Independent verification is still open. Meta’s own report is the main evaluation detail so far; outside SWE-bench Verified or public harness reruns have not yet landed. That is now the normal gap for every new coding model — vendor numbers first, independent numbers after.
Bottom Line
Muse Code and Muse Spark 1.2 are a paired release. The model refines 1.1 for coding; the agent adds the runtime piece for long jobs — persistent agents, isolated subagents, and a crash-safe log — with three small planning skills on top. If the durability holds on real repos, Meta has a plausible third terminal agent. The near-term question is whether teams trade data for the 95% output discount, and whether the long-horizon training translates into fewer steering prompts where these tools are judged.
Sources
- Meta Research - Introducing Muse Code and Muse Spark 1.2 (August 5, 2026)
- Meta - Introducing Muse Spark (April 8, 2026, updated May 12, 2026)
- Reuters - Meta launches new AI coding tool powered by Muse Spark 1.2 (August 5, 2026)
- Wall Street Journal - Meta Releases Coding Agent to Compete With OpenAI and Anthropic (August 5, 2026)
- Wikipedia - Muse Spark (stable release table)
- FutureTools - Meta Launches Muse Code Beta summary (August 5, 2026)
- gHacks - Meta Launches Muse Spark 1.1 in US Public Preview (July 10, 2026)
- TestingCatalog - Meta unveils Muse Spark AI model (April 8, 2026)