Muse Code exits beta with SDK and subscriptions

Meta has formally launched Muse Code with $5, $15 and $50 monthly plans, placing fixed subscription pricing alongside its existing usage-based options. The terminal-based coding agent uses Meta’s Muse Spark 1.2 model, The New Stack reports, after a beta debut less than a month earlier.

At a glance

• The public release adds inter-session messaging, so separate sessions can share context and state when their work overlaps instead of treating related tasks as isolated conversations.
• Workflows coordinate specialized agents over larger jobs, preserve intermediate work between stages, return one final result, and provide a control room where developers can monitor and steer the team.
• The developer-preview SDK lets external software control Muse Code programmatically, including starting or resuming sessions, sending instructions, handling responses, connecting custom tools and streaming the agent’s progress.

Pricing is becoming a central point of differentiation among terminal coding agents, where a low subscription price does not by itself establish how much completed engineering work a plan buys. Meta’s proportional tiers appear intended to make allowances easier to understand, while the continuing token option likely keeps a path for usage patterns that do not fit monthly plans as neatly.

Muse Code adds Workflows, rewind and inter-session messaging

Powered by Meta’s Muse Spark 1.2 model, Muse Code is a terminal-based coding agent positioned against Anthropic’s Claude Code and OpenAI’s Codex. Its beta version used asynchronous background agents to gather information and support the primary agent during a coding session, while Meta presented it for long-running work across large repositories.

The launch adds inter-session messaging, which lets separate Muse Code sessions share context and state when their work overlaps. Workflows orchestrates specialized agents across larger jobs, carries intermediate work between stages and returns a single result. Developers can monitor and steer them through a dedicated control room. A rewind feature rolls a conversation and code changes back to an earlier point.

The $5 plan is estimated at 10 to 50 requests every five hours

Everyday Usage costs $5 per month. High Usage costs $15 and carries three times the usage, while Power Usage costs $50 and carries 10 times the usage. Meta estimates that Everyday Usage typically supports between 10 and 50 requests in each five-hour period, depending on task complexity.

Meta retained its pay-as-you-go rates: Standard charges $1.25 per million input tokens and $4.25 per million output tokens. Contributor charges $0.10 for input and $0.20 for output, but requires users to permit Meta to use prompts and completions to improve its products.

Anthropic includes Claude Code in its $20 Pro plan, with $100 Max 5x and $200 Max 20x plans offering five and 20 times Pro usage and applying five-hour and weekly limits. For GPT-5.6 Sol, OpenAI estimates 10–100 local Codex messages per five hours on $20 Plus, 50–500 on $100 Pro 5x, and 200–2,000 on $200 Pro 20x.

Meta opens Muse Code through an SDK and session protocol

Meta’s new SDK, currently in developer preview, moves Muse Code beyond the command line. It lets software start and resume sessions, send instructions, process responses and react to agent activity, while also supporting custom tools and streamed progress; Zuckerberg said Meta intends developers to build their own agents on top of it.

Meta also published the Muse Session Protocol, which external clients use to communicate with sessions. Developers can use Meta’s TypeScript SDK or implement the protocol directly in another language; Anthropic and OpenAI also offer SDKs that let developers build applications around their coding agents.

The Contributor trade-off Meta has kept pay-as-you-go pricing alongside the new subscriptions, so developers can choose between a monthly allowance and token-based billing. The remaining comparison across coding agents will depend on task complexity, context, reasoning and tools, conditions that Meta and OpenAI both identify as affecting actual usage.