Meta Muse Code Takes On Claude Code and Codex With Cheap Tiers

Meta enters the terminal-agent market

Meta now sells its own AI coding agent. Muse Code is a terminal tool from Meta Superintelligence Labs, the group led by chief AI officer Alexandr Wang, and the company is pitching it directly against Anthropic's Claude Code and OpenAI's Codex.2 It launched on August 5, 2026 as a preview running on Muse Spark 1.2, a new model, and Wang said Meta was competing mainly on price rather than claiming better capabilities.2 Two months on, the product has moved quickly. Its product site says Muse Code left beta on August 31, gained monthly plans, inter-session messaging, multi-agent Workflows and an SDK in developer preview, and switched to Muse Spark 1.3 on September 2.1

The timing explains why this matters now. Muse Code is not entering a new category. It is entering one that Claude Code and Codex already lead and that both companies update almost daily. Meta's chances depend less on whether the agent works, which reviewers generally say it does, and more on whether its pricing and design are enough to pull developers away from tools they already use.

What Muse Code actually is

Coverage agrees on the basics. Muse Code installs with a single command, runs in the terminal, and handles whole engineering tasks across large repositories: it plans a change, writes the code and checks the result.38 The Register called it an agent orchestrator for the command line. When a task starts, background agents keep a context file that working sub-agents can check if they lose track, every action is logged before it runs, and parallel agents work in separate git worktrees.8

Meta's own technical post points to two design choices. The first is the event log: every model call, tool run, approval and edit is written to a local log.3 The second, and probably the more important, is that Muse Spark 1.2 was co-trained with the Muse Code harness. Meta trained it on rejection-sampled harness trajectories and tuned how it handles goals, context compaction and subagents.3 Most coding agents put a general-purpose model inside a wrapper. Meta built the model and the wrapper together. A launch guide noted that this reverses the usual "great general model, wrapper on top" approach.5

The tool also includes review agents. Forbes described implementation running in parallel while reviewer agents watch in the background.4 A September technical breakdown counted four observer agents switched on by default, covering memory recall, skill recommendations, goal tracking and checking that claimed work was actually done. They suggest actions but do not carry them out.7 The same breakdown said Muse Code reads existing AGENTS.md and CLAUDE.md instruction files.7 That feature looks designed to make switching from Claude Code easy.

The pricing play, and the data behind it

Pricing is where Meta is trying to stand out. Wang described a "contributor tier" that he said costs more than ten times less than pay-as-you-go, in exchange for developers agreeing to let Meta use their sessions to improve the model.2 Forbes reported the figures: the standard tier costs $1.25 per million input tokens and $4.25 per million output tokens, while the contributor tier costs $0.10 and $0.20.4 The product site now lists $5, $15 and $50 monthly plans.1

Forbes argued that the discount is really a way to collect data. Real coding sessions are scarce, and they are more useful than scraped or synthetic data because the results can be checked automatically.4 That reading looks right. Meta has always paid for its services with user data, and the contributor tier applies that model to developers. Companies with strict rules about where proprietary code can go will likely be limited to the standard tier, which removes much of the price advantage.

The low prices also come at a hard moment for Meta's finances. CNBC noted that Meta's shares fell the week before launch after a weak revenue forecast and shrinking free cash flow, and that the company is spending heavily on data centers.2 A paid developer product is one of the more direct ways Meta can earn money from that spending.

Benchmarks: where the reporting diverges

The coverage disagrees most on how good Muse Code is. Meta's own launch charts did not show it as the leader. One detailed analysis found that Muse Spark 1.2 trailed Claude Opus 5 in Claude Code on every benchmark Meta published, though it slightly beat GPT-5.6 Terra in Codex and Grok 4.5 on Terminal-Bench 2.1.6 MindStudio was more cautious, calling the debut competitive but not state of the art.9 The Register described Zuckerberg's comparison chart as "somewhat vague."8

The picture changed with Muse Spark 1.3. The same analysis, updated on September 3, said the new model leads or ties Claude Opus 5 and GPT-5.6 Sol on DeepSWE v1.1 and Terminal-Bench 2.1.6 These are still scores Meta chose to publish, and its competitors have kept moving. A weekly roundup says Codex's default model is now GPT-6.1 Sol.25 An unverified post on X claims Anthropic is testing a model called "Claude Next" inside Claude Code.30 Our reading is that Meta has closed the gap to roughly even in its own tests. That is impressive for a two-month-old product, but it gives developers no clear performance reason to switch.

Some details also conflict. The product site says Muse Code does not run on Windows and supports only macOS and Linux.1 A September 20 technical breakdown lists Windows support and gives a PowerShell install command.7 Windows support may have arrived between those two updates, but the conflict is unresolved.

Claude Code's answer: constant, defensive updates

Anthropic has not responded with a big launch. It has kept shipping updates at a very fast pace. Changelog trackers list releases 2.1.286 through 2.1.296 between September 30 and October 9 alone.28 On October 1, version 2.1.287 added Claude Mods, which let plugins change deeper behavior, plus a built-in "You should know" side agent that flags things the user or Claude might miss.2622 Version 2.1.293 made Claude Haiku 5.5 the default Haiku model.21 On the API, Haiku 5.5 has a 1M-token context and short prompts cost a tenth of Haiku 4.5's price.25

Much of this month's work was security work. Releases 2.1.289 to 2.1.292 fixed permission rules that were being skipped, including a security-labeled bypass involving UNC-path file reads. The roundup advises users to update to 2.1.296 or later.25 One changelog tracker counted dozens of fixes in individual releases, several of which quietly closed hook and permission gaps.28 Bugs like these are a reminder of the risk in letting agents run commands on their own, for Claude Code and for every competitor that copies its design.

Code review becomes the battleground

Code review is where these products are competing hardest. Claude Code's recent releases include a steady stream of Code Review fixes. Blocking comments were sometimes labeled "nit" despite their severity, and a pull request's review check could stay "in progress" after the review had finished.21 /code-review output in cloud and SDK sessions now prints as a numbered list instead of raw JSON, and /ultrareview now refuses with an explanation instead of silently leaving uncommitted work out of its upload.21

OpenAI has treated review as a core Codex feature for more than a year. When GPT-5-Codex launched, OpenAI said it had been trained specifically to review code. It compares a pull request's stated intent with the actual diff, reasons across the whole codebase and runs tests, and it can review automatically or when someone types "@codex review."13 Codex has since added a dedicated Codex Security agent14, and one guide cites OpenAI's claim that pre-screening pull requests cuts manual review time by about 60%.15 OpenAI has also gone into Anthropic's own tool: a Codex plugin for Claude Code offers a standard review, a more skeptical adversarial review, and a hand-off of work to Codex.20

Meta's approach is to build review into the agent rather than sell it as a separate pull-request service. Reviewer and verification agents run alongside the agents doing the work.47 That fits a terminal-first product, but none of the reporting describes a GitHub review bot like Codex's or Claude Code's. Teams that want automatic review on every pull request will find Meta's tool lacking for now.

The verdict

Muse Code is a credible product and a serious entry, and it is not yet a threat to the leaders. Its strongest features are co-training the model with the harness, a crash-proof event log, and isolated worktrees. One hands-on review called the worktree sandboxing and crash recovery real differences from rivals, while noting that subagents working on overlapping code still create conflicts someone must fix by hand.10 Codex already has huge reach: more than two million weekly active users by March14 and access bundled into ChatGPT plans.14 Claude Code has a large and still-expanding ecosystem of plugins, hooks, mods and review tools.

Meta's real contribution is putting pressure on price. A contributor tier at a fraction of standard rates4 forces the market to ask what agent-driven coding should cost, and whether developers will trade their sessions for discounts. If enough individual developers accept that deal, Meta gets the training data it is clearly after, and Anthropic and OpenAI face a competitor willing to give away the product to get it.