Meta Muse Code Uses Muse Spark 1.2 to Compete With Claude Code

Meta's coding agent has been out for two months. Has it changed the market?

Meta's first coding agent has been on the market for about two months. In that time the competition has kept shipping updates, and the bigger fight has moved to reviewing AI-generated code. On August 5, 2026, Meta released Muse Code as a beta. It is a terminal coding agent that runs on a new model, Muse Spark 1.2, which Meta calls a step toward the frontier and says will be followed by larger models.1 Engadget described it as Meta's answer to Anthropic's Claude Code and OpenAI's Codex.10 MindStudio put it in the same group of command-line agents as Claude Code, Codex CLI and Grok CLI.3

The main conclusion from the coverage is that Muse Code is a credible and cheap option that does not lead on performance. Its most interesting ideas are about letting developers audit what the agent did, and those ideas line up closely with what Anthropic and the code-review vendors have been building since the launch.

What Meta shipped

Muse Spark 1.2 is a coding-focused update to Muse Spark 1.1. Meta says it improves code generation, debugging, understanding of large codebases and end-to-end developer workflows. Meta also says it used much more training compute on coding tasks for this version.1 The company's developer blog called it a "moderate improvement" aimed at the work agents get most often: refactors across many files, long debugging sessions and tasks that last longer than one prompt.2

The key technical choice is that Meta trained the model and the agent together. Muse Spark 1.2 was trained alongside Muse Code using filtered examples of the agent at work, plus tuning for goals, context compaction and subagents.1 Meta also used a self-improvement loop. Muse Spark 1.1 wrote hard coding environments and instruction templates, then graded candidate solutions, and that produced training data for the new model.4 Build Fast with AI pointed out that Anthropic and xAI use the same kind of tight model-plus-agent pairing with Claude Code and Grok Build.5

On the agent side, Muse Code keeps a set of background agents running for the whole session instead of starting new ones for each task. It also keeps a local, append-only log of every model call, tool run, approval and edit. Meta says the log is "replay-exact" and "restart-safe."4 The developer blog said every subagent, tool call, steer and cancel can be observed and replayed.2 Muse Code runs on macOS and Linux. The model is hosted, and the launch post mentioned no downloadable weights.4 Labellerr reported that the weights are closed and the context window is about 1 million tokens.6

Price is the main pitch

Coverage agreed that cost is how Meta is trying to stand out. By default, Muse Code uses Muse Spark's pay-as-you-go rates of $1.25 per million input tokens and $4.25 per million output tokens. Meta's chief AI officer, Alexander Wang, told CNBC there would also be a much cheaper "contributor" tier.10 That tier has a catch. Meta's documentation says it is limited by tokens over a rolling five-hour window and that its use "may be used to improve our products." The standard model ID costs normal API rates.2 Build Fast with AI put it bluntly: the cheapest tier is cheap because you pay with your data.5

That tradeoff matters to companies. Many AI code review vendors now sell the opposite promise. CodeTuna, for example, says it never stores client code or uses it to train models, and it offers air-gapped deployments.17 Teams with strict data rules will likely see Meta's contributor tier as something for side projects, not their main codebases.

The benchmarks: a solid second tier

Meta's own charts are where the coverage starts to differ. On Terminal-Bench 2.1, Muse Spark 1.2 running in Muse Code scored 82.9%. That beat GPT-5.6 Terra in Codex (81.8%) and Grok 4.5 in Grok Build (81.6%), but trailed Claude Opus 5 in Claude Code (86.7%).9 On DeepSWE 1.1 it scored 59.3%, behind Opus 5 at 65.0% and GPT-5.6 Terra at 64.8%.9

MindStudio reported the same DeepSWE score but placed it behind "GPT 5.6 Turbo" and Opus 5. It also noted that Meta's materials did not compare the model with what it called the current leaders, GPT 5.6 Sol and Fable 5.3 The outside picture is weaker still. Labellerr reported that Vals AI tested the model on a common harness and ranked it 14th of 50 models for general coding.6

Our reading is that the gap between Meta's in-house numbers and Vals AI's ranking is the most important fact in the benchmark coverage. Meta trained the model with its own agent, so scores inside that agent are probably the best case. Build Fast with AI's verdict, that this is "a strong value option, not the new performance king," holds up.5

Meta is moving fast

Verdict, writing via Yahoo Tech, said the release came about a month after Meta opened Muse Spark 1.1 to developers. It also said the Muse Spark series first launched in April and now powers features in the Meta AI assistant.8 Build Fast with AI counted Muse Spark 1.2 as Meta's third model release in about four months.5 Meta also made the model available on OpenRouter and widened global access to its public preview on launch day.2 That looks like a company trying to win developers through reach rather than locking them into its own tools.

Claude Code has kept shipping

Anthropic, the benchmark leader, has not slowed down. On October 1, Claude Code 2.1.287 introduced Claude Mods, which let plugins change deeper behavior. It also added an opt-in built-in mod called "You should know," where a side agent flags things the user or Claude might miss.23 Anthropic describes mods as small TypeScript functions that can rewrite prompts, add interface elements or replace built-in features.22 The October 6 release added an effort setting for sub-agents and direct plugin installs from marketplaces.25 One changelog tracker counted six sandbox and permission gaps closed in that single release.29 By October 7, version 2.1.293 had made Claude Haiku 5.5 the default Haiku model and fixed problems across subagents, Slack and Code Review.21

The two companies are working on similar problems. Muse Code relies on persistent background subagents, and much of Claude Code's October work went into subagent controls, effort settings and making sessions recover more reliably.423 One 2.1.293 fix addressed Claude sometimes redoing or undoing finished work after context compaction.21 That is the same kind of long-session failure Meta says it trained Muse Spark 1.2 to avoid.1

Code review is the next fight

Muse Code's replayable event log matters most when you look at code review. As agents write more code, the hard part has moved to checking it. When Anthropic launched its managed Code Review in March 2026, Boris Cherny said code output per Anthropic engineer was up 200% and reviews had become the bottleneck.13 That managed review costs an average of $15 to $25 per review in research preview, and it never approves or blocks a pull request.13 Even small fixes show how much attention review gets. Version 2.1.293 fixed blocking review comments that were sometimes wrongly labeled as "nits."21 JetBrains researchers noted that Claude Code runs several reviewer agents that tag findings by severity.19

The rest of the review market is changing quickly. GitHub launched ReviewBench, an open benchmark for code review agents built on real pull requests, on October 5.14 Pricing is moving away from flat seat fees. Since June 1, Copilot code review uses AI credits plus Actions minutes. Greptile moved to $1 per review after 50, and Cursor Bugbot moved Teams customers to an on-demand spending pool.11 One developer's Greptile bill reportedly went from $30 to more than $500 in a month.11

Meta has not released a dedicated review product. Its model page says Muse Spark "writes, reviews and ships code," and it links a cookbook for building an autonomous GitHub bot.7 Our analysis is that Meta's combination of low token prices and a full audit log suits a review layer well, but no shipped Meta review product backs that up yet.

The bottom line

Muse Code puts a deep-pocketed, low-cost competitor into a category led by Anthropic. Without independent results that match Meta's own charts, it is a budget option rather than a threat to Claude Code's lead. Its focus on observable, replayable agent runs may matter more than any benchmark score. That focus fits where the industry's attention has gone since August: making sure humans can trust the code agents write.