Meta entered the coding agent race today, and the pricing page is the part worth reading closely.

Mark Zuckerberg announced Muse Code a few hours ago, a terminal-based coding agent now in public beta for macOS and Linux, powered by a new coding-focused model called Muse Spark 1.2. It plans changes, writes code, and validates its own results across large repositories, with persistent background agents that stay alive through an entire session.

That puts Meta directly against Anthropic's Claude Code and OpenAI's Codex, under Alexandr Wang's Meta Superintelligence Labs. But the headline number everyone is quoting, that Muse Code is more than ten times cheaper than the competition, comes with a condition most coverage is burying. Here is the whole thing.

Muse Code installs with a single command and connects straight to the Meta Model API. Unlike Claude Code and Codex, there is no desktop app at all. It lives entirely in your terminal.

The architecture is the interesting part. Muse Code pairs a main agent loop with persistent asynchronous background agents that stay active across a session, gathering context and deciding on their own when to report back, which avoids re-collecting the same information at every stage of a long task. A local event log records model calls, tool use, approvals, and edits, so work resumes after a crash rather than starting over. Built-in commands let you generate a plan, stress-test that plan, or just point it at a goal and let it run.

Meta trained it for genuinely long-horizon work: whole-repository generation, large projects, and automated research. The case study it is leading with is a run optimizing GPU kernels for Nvidia Hopper hardware, where the system made more than 1,000 tool calls across a session lasting up to 24 hours, writing, compiling, and profiling Triton implementations and revising them against a baseline.

Muse Spark 1.2 itself is a coding-focused update to July's 1.1, with reported gains in code generation, debugging, codebase understanding, and long-running workflows. Meta's own launch charts place it second behind Claude Opus 5 on Terminal-Bench 2.1, DeepSWE v1.1, and Meta's internal coding benchmark. Those are vendor-run numbers with no independent verification yet.

Here is the part that deserves more attention than it is getting.

Standard API pricing for Muse Spark 1.2 is unchanged from 1.1: $1.25 per million input tokens, $0.15 cached, and $4.25 output, with a roughly one-million-token context. Competitive, not remarkable.

The number generating headlines is the new contributor tier at $0.10 input and $0.20 output. That is 12.5 times cheaper on input and 21 times cheaper on output.

What you trade for it is permission for Meta to train future models on your prompts and completions. Your code, your architecture, your debugging sessions, your internal patterns. There is also a hard throughput ceiling: 60 requests per minute against the standard tier's 3,000, so it is not viable for production load even if you are comfortable with the terms.

To be fair to Meta, this is disclosed rather than hidden, and for a solo developer working on open-source or personal projects, it may be a perfectly reasonable trade. But "more than 10x cheaper than Claude Code" is doing a lot of work in the coverage today, and the honest version is that the cheap tier is a data-for-discount swap with a rate limit that rules out most real workloads. If you work anywhere with a client NDA or proprietary code, the standard tier is your only option, and at $1.25 and $4.25 it is competitive rather than disruptive.

Step back and today reads as the moment the coding agent became the product every frontier lab has to ship.

Anthropic has Claude Code. OpenAI has Codex. DeepSeek adapted V4-Flash for Codex compatibility last week specifically so it could run inside OpenAI's tooling. And now Meta has a terminal agent of its own, co-trained with a model built for it.

Notably, Meta chose not to build an app. Claude Code and Codex both have desktop surfaces. Muse Code is terminal-only, which is a bet that the developers worth winning are the ones already living in a shell, and that Meta does not need to fight a UI battle to compete.

There is one more thing worth flagging. Muse Spark remains closed weights, with no repository in Meta's Hugging Face organizations. For a company whose entire AI identity was built on open Llama models, that is now the third closed release in a row, and it puts Meta in the odd position of being less open than Alibaba, Moonshot, and DeepSeek while competing on the same price axis.

If you are already deep in Claude Code or Codex, nothing here forces a switch today. The benchmarks are Meta's own, the model sits second to Opus 5 by Meta's own admission, and there is no app. But the long-horizon architecture is worth a look if you run genuinely large tasks, since persistent background agents plus a crash-resumable event log addresses a real failure mode in agent workflows that most tools handle badly.

The strategic read is simpler. A month ago the coding agent market was two products. Today it is four, everyone is undercutting everyone, and one vendor is willing to charge you almost nothing in exchange for your codebase. That is what a market looks like right before margins disappear. Good for you, and worth watching what it does to the people selling it.

The stories moving fast beyond today's headlines:

Mark Zuckerberg announced Muse Code today, Meta's first AI coding agent, a terminal-only tool in public beta for macOS and Linux powered by a new coding-focused model, Muse Spark 1.2, and built by Alexandr Wang's Meta Superintelligence Labs. It plans changes, writes code, and validates results across large repositories, pairing a main agent loop with persistent asynchronous background agents and a local event log that lets work resume after a crash, with a launch case study running more than 1,000 tool calls over a session lasting up to 24 hours while optimizing Nvidia Hopper GPU kernels. Meta's own charts place Muse Spark 1.2 second behind Claude Opus 5 on Terminal-Bench 2.1, DeepSWE v1.1, and an internal benchmark, all vendor-run and unverified. Standard API pricing is unchanged from 1.1 at $1.25 input, $0.15 cached, and $4.25 output per million tokens with a roughly one-million-token context, while the headline-grabbing contributor tier at $0.10 and $0.20 is 12.5 to 21 times cheaper in exchange for permission to train future Meta models on your prompts and completions, capped at 60 requests per minute against the standard tier's 3,000. That makes the cheap tier a data-for-discount trade unsuitable for proprietary or client work and too rate-limited for production. Muse Spark remains closed weights with no Hugging Face repository, leaving Meta less open than Alibaba, Moonshot, and DeepSeek while competing on the same price axis. The coding agent market went from two products to four in a month, and margins are the thing to watch.

Stay building. 🤖