Meta Muse Code: Pricing, Benchmarks, Data-Sharing Catch

Meta is entering the AI coding wars. On August 5, 2026, the company launched Muse Code, its first terminal-based coding agent, alongside a new model called Muse Spark 1.2 that powers it. The tool is available in public beta for macOS and Linux, and it goes straight after the market that Anthropic’s Claude Code and OpenAI’s Codex have defined over the past year — with an aggressive pricing structure and a data-sharing tier that is the most controversial part of the launch.

The effort is overseen by Alexandr Wang, head of Meta Superintelligence Labs, and it extends the strategy Meta began in July when it started charging for its models for the first time.

What Muse Code does

Muse Code is a terminal agent built for large codebases. It takes on complete software-engineering tasks — planning a change, writing the code, and validating the result — rather than just autocompleting lines. Its headline architectural feature is parallelism: for big jobs, the main agent launches its own sub-agents that work simultaneously, each in an isolated git worktree, before their results are merged back.

That fan-out-to-subagents pattern is the same architecture most serious agent frameworks have converged on in 2026. Muse Code also supports persistent async background agents — long-running tasks that continue working while the developer does other things — which Meta pitches as the differentiator against tools that block on a single foreground session.

The model underneath is Muse Spark 1.2, an update to the Muse Spark 1.1 model Meta shipped in July. It is tuned specifically for the coding-agent workload: long-horizon tool use, repository navigation, and validation loops.

The pricing: a standard tier and a much cheaper “contributor” tier

Muse Code’s pricing is where Meta’s strategy diverges sharply from its rivals. There are two tiers.

The standard pay-as-you-go tier is priced at $1.25 per million input tokens and $4.25 per million output tokens, with cached input at $0.15. On this tier, Meta commits that prompts and completions are not used to train its models — matching the privacy posture developers expect from Anthropic and OpenAI, at a broadly comparable price.

The contributor tier is the eye-catching part: $0.10 per million input tokens and $0.20 per million output tokens, with cached input at a near-free $0.002. That is roughly 12x cheaper on input and 21x cheaper on output than the standard tier — in exchange for developers opting in to let Meta use their code and interactions to improve the model.

In other words, Meta is offering to subsidize your AI coding bill down to almost nothing if you’ll pay in data. It is the same trade that powered Meta’s advertising business, now applied to developer tooling: the product is cheap because you are helping build the next version of it.

The benchmarks, with an asterisk

Meta says Muse Spark 1.2, running inside Muse Code, scored 82.9% on Terminal-Bench 2.1. In Meta’s own comparison, that edges OpenAI’s GPT-5.6 Terra in Codex (81.8%) and xAI’s Grok 4.5 in Grok Build (81.6%), while trailing Anthropic’s Claude Opus 5 at 86.7%.

The methodology deserves scrutiny, and Meta was transparent about it. Rather than running every model through a neutral harness, Meta ran each model inside its own vendor’s agent product — Muse Code for Muse Spark 1.2, Claude Code for Opus, Codex for GPT, Grok Build for Grok — in isolated Daytona cloud sandboxes, scored pass@1 averaged over five attempts across 89 tasks. That is arguably the most realistic way to compare products developers actually use, but it also means the numbers reflect the agent-plus-model combination, not the raw model, and self-reported benchmark results always warrant independent verification.

The takeaway is not that Muse Spark 1.2 is the best coding model — Opus 5 still leads Meta’s own chart — but that Meta has reached the competitive tier and is willing to compete on price rather than on being clearly better.

Meta’s internal proving ground

Meta is dogfooding the tool hard. The company is requiring thousands of engineers to use Muse Code weekly, and it reports roughly 7,000 active internal users who have already generated over 800 fixes that were fed back to improve the model. That internal flywheel — engineers use the tool, their corrections train the next version — is the same dynamic the contributor tier extends to the public.

It also explains the pricing bet. The contributor tier isn’t primarily a revenue play; it’s a data-acquisition play. Meta wants the volume of real-world coding interactions that Anthropic and OpenAI already collect from millions of Claude Code and Codex users, and it’s willing to price aggressively to buy its way into that flow.

The market’s muted reaction

For all the strategic weight of the launch, Meta’s stock was roughly flat in after-hours trading on August 5 — a notable contrast to the two large single-day gains the stock posted in July when Muse Spark 1.1 debuted. Investors appear to have already priced Meta’s AI pivot into the shares, and a follow-on release, however aggressive, didn’t move the needle. The market is now waiting to see whether the spending on this strategy converts into revenue and usage rather than rewarding each announcement.

What it means

Meta is competing on price and data, not superiority. The Muse Spark 1.2 benchmarks put Meta in the game but behind Opus 5. So Meta is doing what it does best: undercutting on price and monetizing attention — here, developer interactions — instead of winning on raw capability. The contributor tier makes AI-assisted coding nearly free for anyone willing to trade their code, which could pull price-sensitive developers, students, and open-source maintainers away from the paid incumbents.

The privacy trade is the real story. Two tiers at a 12–21x price gap force a decision developers haven’t had to make so explicitly before: keep your code private at market rate, or hand it over for near-free access. For hobby projects and public repos, the math favors the cheap tier. For proprietary enterprise code, it’s a non-starter — which is exactly the boundary Anthropic and OpenAI will lean on to defend their business customers.

The coding-assistant market just got a well-funded third entrant. Claude Code and Codex had largely split the serious agentic-coding market between them. Meta arriving with comparable benchmarks, a parallel-subagent architecture, and the deepest pockets in tech changes the competitive math — especially on price, where Meta can afford to lose money longer than a startup can.

Watch the data flywheel, not the launch. The 800 internal fixes and the contributor tier point at the same thing: Meta is trying to bootstrap the real-world coding data that its rivals already have. If the cheap tier attracts enough volume, Muse Spark’s coding ability could close the gap with Opus quickly, because the model improves on exactly the interactions it’s now collecting. The flat stock says investors will believe it when the usage — and the revenue — show up.