Meta spent years watching Anthropic and OpenAI turn AI coding assistants into their most reliable revenue engines. On August 5, 2026, it finally answered back. Muse Code, Meta’s first dedicated AI coding agent, launched in public beta alongside a new coding-focused model called Muse Spark 1.2, and the pitch is blunt: similar agentic capability to Claude Code and OpenAI’s Codex, at a fraction of the price.
The announcement came from CEO Mark Zuckerberg on X, with the technical rollout led by Alexandr Wang, the Scale AI founder who now runs Meta Superintelligence Labs. It landed a week after Meta’s stock dropped on a soft revenue forecast and shrinking free cash flow, which makes the timing hard to ignore. Meta has poured hundreds of billions of dollars into AI infrastructure. Muse Code is the first product built to visibly justify that spending to developers, not just to investors.
This piece breaks down what Muse Code actually does, how its own published benchmarks stack up against Claude Code and Codex, what it costs, and where it currently falls short. The short version: it is not the strongest coding agent on the market by Meta’s own numbers, but the pricing structure is aggressive enough that it changes the calculus for a lot of engineering teams, especially at scale.
What Muse Code Actually Is
Muse Code is a terminal based coding agent, not a chatbot with a code tab bolted on. It runs from the command line, reads and reasons across large repositories, plans multi step changes, writes code, executes tests, and validates the results before handing work back to a developer. According to Meta’s own research post, the agent can also coordinate persistent background subagents that keep working on separate parts of a task while the main session continues, and it maintains a local event log of every model call, tool run, approval, and file edit.
That event log matters more than it sounds. Long running agent sessions fail in mundane ways: a laptop sleeps, a terminal window closes, a network drops mid task. Because Muse Code records state continuously, a session can resume exactly where it stopped instead of re deriving context from scratch. For a fifty call refactor that’s a nice convenience. For a multi hour job involving hundreds or thousands of tool calls, which is increasingly how these agents get used for things like kernel optimization or large scale migrations, it’s closer to a requirement.
Muse Code ships for macOS and Linux, installed through a one line shell command, with no dedicated desktop app at launch. That puts it a step behind Claude Code and Codex, both of which have broader interface options including IDE integrations. Meta says more harness features are coming, and the product is explicitly labeled a beta.
# Example install pattern Meta describes for Muse Code (verify current command before running)
curl -fsSL https://ai.meta.com/muse-code/install.sh | shThe Model Underneath: Muse Spark 1.2
Muse Code runs on Muse Spark 1.2, a coding focused update to July’s Muse Spark 1.1. Meta says the two were co-trained together, meaning the model and the agent harness were developed in tandem rather than a general purpose model being wrapped in a generic agent loop afterward. Meta’s argument is that this pairing produces more reliable tool use and fewer wasted retries. It’s a reasonable theory, and one that’s genuinely hard for outsiders to verify independently at this stage.
The model carries a 1 million token context window, inherited from the Muse Spark line, which is large enough to hold substantial portions of a real codebase in a single session. Meta says training compute on coding tasks was scaled up significantly for 1.2, along with a wider range of training environments, while the model retains the general agentic strengths (tool use, multi step planning, computer control) that defined Muse Spark 1.1.
One detail that got buried under the bigger headlines: Muse Spark 1.1 was US only at launch. Muse Spark 1.2 ships with expanded global access through the Meta Model API. For developers outside the United States, that access change is arguably more consequential than any individual benchmark score.
How It Compares to Claude Code and Codex
Meta published its own benchmark comparisons at launch, running Muse Spark 1.2 inside Muse Code against Claude Opus 5 inside Claude Code and GPT-5.6 Terra inside Codex. A caveat worth stating plainly: every one of these numbers is vendor run, each model was tested inside its own harness, and there is no neutral third party audit of the results yet. Treat the table as Meta’s self reported starting point, not settled fact.
Benchmark Muse Code (Muse Spark 1.2) Claude Code (Claude Opus 5) Codex (GPT-5.6 Terra) Terminal-Bench 2.1 82.9% 86.7% 81.8% DeepSWE 1.1 (Meta internal) 70.6% 79.4% Not published in Meta’s chart Long tool-call kernel task (1,000+ calls) ~61 to 69% improvement ~74 to 75% improvement Not published in Meta’s chart
Read plainly, Meta’s own data shows Muse Code trailing Claude Code on the coding heavy benchmarks it chose to publish, and running roughly even with or slightly ahead of Codex depending on the test. That is not the story Meta led with in its marketing, but it’s the story its own numbers tell.
Where Meta does have a case is durability under load. In the extended kernel optimization test, which involved more than a thousand sequential tool calls without human intervention, the gap between Muse Code and Claude Code narrowed compared to shorter benchmark runs. A workload that runs unattended for nine hours has different failure modes than a fifty step refactor, and Meta’s replay exact session logging is built specifically for that kind of long horizon reliability. Whether that translates into fewer real world failures than Claude Code’s own long running session handling is not something the published benchmarks settle either way.
It’s also worth remembering the backdrop. Anthropic’s Claude Code became one of the company’s largest revenue drivers after its 2025 launch and effectively created the category Meta is now entering. OpenAI followed with Codex. GitHub Copilot, backed by Microsoft, still holds the largest share of the day to day developer market even as these agentic tools grow. Muse Code isn’t walking into an empty field. It’s the fourth serious entrant into a market that already has an established leader, a fast follower, and an incumbent with enormous distribution.
Pricing: Where Muse Code Actually Differentiates Itself
Benchmarks aside, price is the part of this launch that developers are most likely to act on. Muse Code has no subscription tier. Everything is billed by tokens, on one of two plans.
Tier Input (per 1M tokens) Cached input (per 1M tokens) Output (per 1M tokens) Data usage Standard $1.25 $0.15 $4.25 Not used to train Meta’s models Contributor $0.10 $0.002 $0.20 May be used to improve Meta’s models
The Contributor tier is the aggressive part of this launch. It runs roughly 12 times cheaper on input tokens and about 21 times cheaper on output tokens than the Standard tier, in exchange for letting Meta use submitted prompts and code to improve its models. Wang described it publicly as “more than ten times cheaper” than comparable offerings elsewhere. It’s currently limited to select countries and runs on a rolling five hour rate limit rather than the flat per minute limits on the Standard tier.
For a team that’s comfortable with the data sharing trade off, and whose code isn’t proprietary in a way that makes that uncomfortable, the Contributor tier meaningfully undercuts both Claude Code and Codex on raw per token cost. For teams handling client work, regulated data, or anything under an NDA, the Standard tier (or a zero data retention arrangement, which Meta says it is now rolling out for enterprise customers) is the only realistic option, and at that price point the cost advantage over competitors narrows considerably.
Meta also listed Muse Spark 1.2 on OpenRouter, the multi model API aggregator that already hosts models from labs including DeepSeek and Z.ai. That makes it easy for teams already using an aggregator to A/B test Muse Code against Claude and GPT based agents without a separate procurement process.
The Real Benefits
Cost at scale. For high volume, non sensitive workloads, the Contributor tier pricing is genuinely disruptive. Teams running thousands of automated coding tasks a day, think CI adjacent code review, test generation, or routine refactors, stand to save real money compared to Claude Code or Codex at list price.
Long context out of the box. The 1 million token window means large codebases can be loaded into a single session without aggressive chunking, which reduces the engineering overhead of managing context manually.
Session durability. The replay exact event log is a genuinely useful design choice for long running, unattended agent work, and it’s a feature Meta built specifically around real failure modes rather than benchmark optics.
Expanded global access. Developers outside the US, previously locked out of Muse Spark 1.1, now have a path in, which widens the addressable market for anyone building on Meta’s models.
The Costs and Trade-offs
Single provider risk. Muse Spark 1.2 currently runs on Meta’s infrastructure alone. Claude and GPT models are available through multiple cloud providers with automatic failover options; Muse Spark 1.2 is not, at least not yet. That’s a structural risk for any team building production workflows on top of it.
The data trade on the cheap tier. The Contributor tier’s discount exists because Meta gets to use the data. For a company whose entire business model runs on turning user data into targeting and product improvements, that trade deserves more scrutiny than a passing mention in a pricing table, particularly for anyone pointing it at proprietary or client owned code.
No unified app experience yet. With no dedicated desktop or IDE native app at launch, Muse Code asks more of developers up front than tools that already live inside an editor.
The Limitations
The clearest limitation is the one Meta’s own charts show: on the coding specific benchmarks it chose to publish, Muse Code currently trails Claude Code, sometimes by a meaningful margin. Meta is not hiding this, its evaluation report is public, but it does mean the “beats Anthropic and OpenAI” framing in some coverage overstates what the company’s own data supports today.
Independent, third party verification is also still missing. Every comparison published so far, including the ones in this article, comes from Meta running its own model inside its own harness against competitors it configured itself. That is standard practice across the industry at launch, but it means real world developer testing over the next few months will matter more than the launch benchmarks.
Regional and rate limits on the Contributor tier reduce its practical value for teams outside the eligible countries, and the beta label is not a formality. Meta has already flagged additional harness features as still in development.
Should You Actually Switch?
The honest answer depends on what kind of team you run, not on which company’s marketing you trust more.
Stick with Claude Code if your workloads are complex, high stakes, or already tuned around Anthropic’s harness and tool ecosystem. Meta’s own Terminal-Bench and DeepSWE numbers put Claude Code ahead on the exact tests Meta chose to publish, and switching agents mid project has real transition costs that a lower per token price won’t offset for a small team.
Consider Codex if you’re already deep in the OpenAI ecosystem, or need an agent with the broadest existing IDE and plugin support. Muse Code’s Terminal-Bench score sits close to Codex’s, so the deciding factor there is more likely to be tooling fit than raw capability.
Consider testing Muse Code if you run high volume, low sensitivity workloads (internal tooling, test generation, routine dependency upgrades) where the Contributor tier’s data sharing trade off is acceptable and the cost savings compound quickly across thousands of daily calls. It’s also worth a look if your team is outside the US and was previously locked out of Muse Spark 1.1, since the expanded access is new as of this release.
Wait if your code is proprietary, client owned, or under an NDA, and you’re not comfortable with the Standard tier’s four times higher output cost. In that case Muse Code’s price advantage over Claude Code and Codex shrinks enough that the benchmark gap becomes the more important factor.
Because Muse Spark 1.2 is listed on OpenRouter alongside other frontier models, running a real side by side test on your own repository, rather than trusting any single company’s launch chart, is now a matter of an API key swap rather than a procurement process. That’s probably the most useful thing this launch actually delivers to developers: a cheap, low friction way to check Meta’s claims against your own codebase.
Why Meta Is Doing This Now
None of this is happening in a vacuum. Meta gets roughly 98% of its revenue from advertising, a business built entirely around user data, and it has spent enormous sums building AI infrastructure with comparatively little to show shareholders in direct product revenue. Coding tools are one of the few AI categories with a proven willingness to pay: Anthropic’s Claude Code helped turn the company into a genuine enterprise revenue source, and that example is not lost on Meta’s leadership.
Alexandr Wang, who joined Meta from Scale AI roughly a year ago specifically to rebuild its foundation model efforts, has now shipped three Muse Spark releases in four months. Muse Code is the first of those releases aimed squarely at generating direct revenue rather than powering Meta’s own consumer apps. Whether it succeeds depends less on the launch benchmarks than on what happens over the next two or three release cycles, and on whether Meta can close the gap its own charts currently show against Claude Code.
What to Watch Next
Muse Code is a credible first entry, not a finished argument. The pricing is the most interesting part of this launch by a clear margin, and the session durability design shows real product thinking. But developers evaluating it seriously should run their own tests rather than take any vendor’s launch benchmarks, Meta’s included, at face value. Watch for independent SWE-bench and Terminal-Bench runs over the coming weeks, watch whether the Contributor tier’s country restrictions loosen, and watch how quickly Meta ships the next Muse Spark update. In a market this new, this launch is a data point, not a verdict.