Meta Priced a Coding Agent at $0.30 and Named the Currency: Your Code

Meta shipped a terminal coding agent on August 5. It finishes third.

I know that because Meta published the charts showing it finishing third. On Terminal-Bench 2.1, Muse Spark 1.2 scores 82.9% against Claude Opus 5’s 86.7%. On DeepSWE 1.1, it’s 59.3% against Opus 5’s 65.0%. On Meta’s own internal benchmark, built from 440 real pull requests out of Meta’s codebase, it’s 70.6% against 79.4%.

Three charts, Meta’s choice of opponents, Meta’s own harness. Claude wins all three.

Which makes the launch far more interesting than a win would have been.

The Number That Is Actually the Product

Standard tier: $1.25 per million input tokens, $4.25 per million output. Competitive, unremarkable.

Contributor tier: $0.10 input, $0.20 output. That is 12.5x cheaper on input and 21x cheaper on output. Blended, it lands around $0.30 per million tokens, which makes it the cheapest coding model on the market, undercutting the previous floor by a comfortable margin.

For scale: Claude Opus 5 runs $5 input and $25 output. Blended, roughly $30. The Contributor tier is on the order of 100x cheaper.

Meta’s Chief AI Officer Alexandr Wang said the quiet part into a CNBC microphone: Meta is differentiating “by price rather than capabilities.” That is a remarkably honest thing to say on launch day, and it tells you exactly what this product is. Muse Code is not a bid to build the best coding agent. It is a bid to make the best coding agent’s price look indefensible.

And the target is worth the shot. Claude Code reached $1 billion in annualized revenue in six months and $2.5 billion in nine, writes roughly 4% of all public GitHub commits, and has 1,000+ customers spending over $1 million a year — a figure that doubled from 500+ in under two months earlier this year. OpenAI has disclosed 5 million+ weekly Codex users. That is a market large enough that undercutting it by 100x is a rational strategy even if you never win a single benchmark.

What the Discount Costs

The Contributor tier is not a promotional rate. It is a transaction.

Standard tier, per Meta: prompts and completions are not used for training. Contributor tier: you grant explicit permission for Meta to store, review, and train on your prompts and your code. The 21x discount is paid in training data.

The rate limits confirm who each tier is for. Standard allows 3,000 requests and 4 million tokens per minute per team. Contributor allows 60 requests per minute — a 50x tighter ceiling. That is not a production budget. That is an individual developer working on a side project, which is precisely the population most likely to accept the trade and least likely to have a lawyer weigh in.

There is a genuine strategic question buried in this, and it cuts against Meta. If the developers who accept the deal are hobbyists and open-source contributors, and the enterprises with the interesting proprietary code all stay on Standard or leave for Anthropic, then Meta is buying a training corpus of exactly the code that was already public. The flywheel spins, but it may be spinning on material Meta could have scraped anyway.

Wang also noted Meta is “starting to accept requests for zero-data retention.” Requests. Not a setting.

The Part Meta Deserves Credit For

The model is third. The harness is genuinely good, and several developers reacting on launch day landed on the same read: the harness is the story, not the model.

Three bets are worth understanding even if you never install this thing:

Persistent async background agents. Rather than spawning a fresh agent per task, agents stay alive across the whole session and accumulate context. Meta demonstrated this with a kernel optimization run: 1,000+ tool calls over as long as 24 hours on NVIDIA Hopper hardware, optimizing Triton kernels with third-party kernel libraries prohibited. Zuckerberg’s claim is that “it kept finding substantial improvements well beyond the initial exploration phase.”

That claim, if it survives contact with other people’s code, addresses the most common real complaint about coding agents: they don’t fail loudly, they plateau quietly.

Parallel sub-agents in isolated git worktrees. Each sub-agent gets its own worktree, so your working copy is never the shared mutable state. Zuckerberg says they had it build six features for a game simultaneously with no collisions. Anyone who has watched two agents fight over the same file understands why this is the right primitive.

An append-only local event log. Every model call, tool run, approval, and edit is written to disk before it executes. The runtime is replay-exact and restart-safe after a crash. This is the least glamorous feature and the most professional one. It is what infrastructure looks like when someone expects it to fail and plans for the recovery instead of the demo.

Where the Benchmarks Get Slippery

One disclosure in Meta’s own material deserves more attention than it got.

Meta notes that the Muse Spark 1.1 comparison scores were run in a generic harness, while 1.2 was run inside Muse Code. So the headline generational gain — roughly +6.7 on Terminal-Bench, +6.3 on DeepSWE — is split between the model and the harness in an undisclosed ratio. Muse Spark 1.2 was also co-trained on rejection-sampled trajectories from the Muse Code harness itself, which makes the model and the scaffolding genuinely hard to separate.

Independent measurement narrows the numbers further. Artificial Analysis, with pre-release access, measured Terminal-Bench v2.1 at 80%, not 82.9%, and put Muse Spark 1.2 at 54 on its Intelligence Index — up 3 points from 1.1, tied with Grok 4.5, behind Claude Opus 5 at 61. Reddit commenters noted it matches Grok 4.5’s score while burning roughly 50% more tokens, which quietly erodes some of the price advantage on real tasks.

The Llama Reversal

Muse Code is fully closed. No open weights, no license, no self-hosting. Neither the launch blog post nor Zuckerberg’s thread uses the word “open.”

This is the company whose Llama family reached roughly 1.2 billion downloads, and whose CEO published an essay in July 2024 titled “Open Source AI Is the Path Forward.” Asked directly on X whether Muse Code would be open source, Zuckerberg replied only: “I’ll have more to share on that soon.”

The irony is sharp when you check the neighbors. OpenAI’s Codex CLI is Apache 2.0. Google’s Gemini CLI is open source. On this particular axis, Meta has moved to the most proprietary position of the three.

There is also a hard availability floor worth knowing before you plan around it. macOS and Linux only — Windows developers are on WSL or nothing, which is a fleet-level constraint for a lot of enterprises, not an inconvenience. No GUI, no IDE integration, and the launch material does not mention MCP support. VentureBeat’s own install test got the one-line installer working, then hit a wall: no models visible and payment “required to finish setting up your account.” Low-cost is accurate. Free is not.

The Business Takeaway

For most teams, the correct move this week is not migration. It is writing down a policy.

Separate your repositories by confidentiality before anyone benchmarks anything. The Contributor tier is a reasonable deal for code you’d be comfortable publishing, and a serious problem for anything under an NDA, client confidentiality, regulatory obligation, or trade-secret protection. That is a decision that belongs in writing, made once, at the organization level — not made ad hoc by whichever developer noticed the discount first.

Price the whole task, not the token. A model that costs 100x less per token but uses 50% more tokens and requires more correction passes is not 100x cheaper. Run your own repository through it and measure the cost-to-merged-PR, because that is the only number that hits your budget.

Treat the harness as the upgradeable part. The most durable lesson here isn’t about Meta. It’s that a meaningful share of a coding agent’s performance lives in the scaffolding — the persistence, the isolation, the crash recovery — not the weights. That means your own agent setup has headroom you can capture without switching models at all.

And note what a three-horse race does to your bill. Meta just established a credible price floor two orders of magnitude below frontier pricing. Whether or not you ever run Muse Code, that floor is now an input to your next renewal conversation.

The Fog

The fog around AI coding agents has been a benchmark fog: a stack of percentages, each measured in a different harness, each chosen by the vendor publishing it. Meta’s launch cuts through some of it precisely by being honest about the trade. It published charts it loses, and a price it wins on, and named the currency.

The clarity is in the trade itself. You can have frontier capability, or you can have a 100x discount funded by your own source code. Both are legitimate choices. What matters is knowing which one you are making, on which repository, on purpose.

AI eliminates the fog. It doesn’t eliminate the decision — it just makes the decision legible enough to make on purpose, in writing, before the first commit lands.

Sources: Meta AI Research — Introducing Muse Code and Muse Spark 1.2, VentureBeat, CNBC, Artificial Analysis, 9to5Mac, Meta Superintelligence Labs background