Muse Code and Muse Spark 1.2: Meta's Coding Agent Offers 20x Cheaper Tokens, If You Give Up Your Data

Meta just entered the coding agent war, and its opening move is a pricing bomb with a privacy catch.

On August 5, 2026, Meta released Muse Code in beta, a terminal coding agent built to go head to head with Anthropic’s Claude Code and OpenAI’s Codex. It runs on Muse Spark 1.2, the company’s latest model and a meaningful step past the Muse Spark 1.1 that launched a month earlier.

For heavy AI users, the headline number is hard to ignore: sign up for Meta’s “Contributor” tier, let Meta train future models on your prompts, and the token price drops from $1.25 in and $4.25 out per million tokens to just $0.10 in and $0.20 out. That is up to 20x cheaper. The tradeoff is your data, and you should weigh it carefully before you hit install.

Here is exactly what Muse Code and Muse Spark 1.2 mean for your costs, your quotas, and your privacy.

Muse Spark 1.2 Pricing: Two Tiers, a 20x Gap

The most surprising thing about Muse Code is not the agent itself. It is that Meta is using a discount-for-data model that reshapes how you should think about its pricing.

Tier	Input $/MTok	Output $/MTok	Rate limit
Standard	$1.25	$4.25	3,000 req/min, 4M tok/min
Contributor	$0.10	$0.20	60 req/min, 2.1M tok/min

The Contributor tier costs 12.5x less on input and 21x less on output. If you are a heavy API user moving serious token volume through agentic coding, that difference is the difference between a $400 monthly bill and a $30 one.

But the rate limits tell the real story. The Contributor tier is capped at 60 requests per minute and 2.1 million tokens per minute, versus 3,000 requests per minute for Standard. So the cheap tier is generous on tokens but restrictive on request bursts. For single-agent coding sessions that is usually fine. For fleets of parallel agents, it becomes a ceiling you will hit fast.

Muse Code vs Claude Code: What Changes

Claude Code has been the default terminal agent for many heavy users, and Codex has been closing the gap. Muse Code attacks both on capability and on price.

The agent ships with a persistent event log where every model call, tool run, approval, and edit is appended. That makes each session replay-exact and restart-safe: after a crash, the agent resumes exactly where it stopped. For long-running tasks this removes a huge source of wasted tokens, because you are not re-sending context after failures.

It also uses persistent background subagents that stay active across a session instead of being spawned and killed per task. That cuts redundant information gathering, which translates directly into fewer tokens spent on re-reading context.

Built-in skills include /plan to turn a task into an approval-gated plan, /grill to stress-test that plan, and /goal to drive toward an objective. For the type of heavy user who lives in the terminal, this is a credible alternative to the Claude Code workflow at a fraction of the unit price.

The Contributor Tier Is a Data Leak You Should Price

Here is the part no pricing page will emphasize. The 20x discount exists because Meta wants your prompts and completions to train future models. That is a genuine tradeoff, not a free lunch.

Give Meta access to your coding sessions and you are feeding your proprietary logic, your architecture decisions, and your internal naming conventions into the training corpus of a competitor’s model. For an individual developer that may be an acceptable bargain. For a company shipping source code, the legal and security exposure is often not worth the savings.

If you are deciding between the two tiers, do the math on what your data is worth, not just what your tokens cost. The Contributor tier is the cheapest coding agent on the market for a reason, and that reason is you.

What Heavy AI Users Should Do Now

First, treat Muse Code as a legitimate third option in the coding-agent race. It is genuinely capable, restart-safe, and priced aggressively enough to force Claude Code and Codex to respond.

Second, if you are cost-sensitive and your projects are not sensitive, the Contributor tier at $0.10 in and $0.20 out is arguably the best deal in agentic coding today. Just be honest about what you are trading.

Third, if your work involves proprietary code or client data, pay the Standard tier or skip Muse Code entirely and stay on a tool whose pricing does not come with a training-data invoice.

The bigger picture matters more: the price war has now reached coding agents, not just raw models. Every provider is racing to the bottom on per-token cost, and the ones that cannot win on efficiency will start asking for your data instead of your dollars. Track your spend closely, because the cheapest option on paper is rarely the cheapest option once you account for the privacy tax.