Meta has moved Muse Code, its terminal-based coding agent, out of beta. The switch comes less than a month after the beta launch, and the headline items are three subscription plans starting at 5 USD a month plus an SDK released as a developer preview. Against Anthropic's Claude Code and OpenAI's Codex, both of which got there first, Meta is competing head-on with price.
Three tiers at 5, 15 and 50 USD
The new plans are Everyday Usage, High Usage and Power Usage. Everyday Usage costs 5 USD a month (about 800 yen), High Usage costs 15 USD a month (about 2,400 yen) and delivers three times the usage of Everyday Usage, and Power Usage costs 50 USD a month (about 8,000 yen) for ten times the usage.
※1 USD = 160 JPY (as of September 1, 2026)
For the entry tier, Meta puts the guideline at roughly 10 to 50 requests every five hours, with the caveat that the number moves with the complexity of the work.
Line the competition up and the gap is clear. Claude Code starts with Pro at 20 USD a month (about 3,200 yen), with Max 5x at 100 USD (about 16,000 yen) and Max 20x at 200 USD (about 32,000 yen). Codex is bundled with the 20 USD Plus plan, and the Pro tiers at 100 USD and 200 USD offer 5x and 20x the usage of Plus. Muse Code's entry point sits at roughly a quarter of what its rivals charge.
Comparing those numbers directly is risky, though. Anthropic does not publish absolute request counts, expressing its allowances as multiples of Pro alongside five-hour and weekly caps. OpenAI cites a range for GPT-5.6 Sol of 10 to 100 messages every five hours on Plus, 50 to 500 on Pro 5x and 200 to 2,000 on Pro 20x, while noting that actual usage shifts with the model, task complexity, context, reasoning and tool use. Meta's 10-to-50-request figure is the same kind of number, so none of this lines up neatly.
Pay-as-you-go, and what makes it cheap
The usage-based pricing from the beta period stays in place alongside the subscriptions. The standard rate is 1.25 USD per million input tokens and 4.25 USD per million output tokens.
Below that sits a Contributor tier, where the rates drop an order of magnitude to 0.10 USD for input and 0.20 USD for output. The condition is that users opt in to having their prompts and completions used to improve Meta's products. For teams handling proprietary code, that rules the tier out immediately. The source of the discount is written directly into the price list.
Skepticism about Muse Code concentrates on exactly this point. Meta's revenue comes from advertising, so the argument goes that unusually low pricing is a way in to collect development data. The debate from the beta period carries over unchanged.
Three new features for wrangling agents
The release also brings functional changes. Underneath it all runs Meta's own Muse Spark 1.2.
Inter-session messaging lets separate Muse Code sessions hand context and state to each other directly when their work overlaps. A warning raised in one terminal no longer has to be retyped into another.
Workflows builds on the multi-agent capability that existed in beta. Specialized agents are organized around a larger job, intermediate work is carried between stages, and a single result comes back at the end. A dedicated control room lets developers watch progress and steer the agents mid-run.
Rewind rolls a session's conversation and code changes back to an earlier point together. For a tool built around agents that run for a long time, being able to choose how far to unwind is itself a safety mechanism.
The SDK takes it outside the CLI
The other headline item is the SDK, released as a developer preview. It allows Muse Code to be driven programmatically: starting and resuming sessions, sending instructions, handling responses and reacting to what the agent does, all without a person sitting at the terminal. That opens the door to IDE integrations, internal engineering tools and standalone agent products built on top of Muse Code.
Underneath is the Muse Session Protocol (MSP), the protocol external clients use to talk to Muse Code sessions. Meta ships a TypeScript SDK, but implementing MSP directly allows connections from other languages.
The direction is not novel in itself. Anthropic's Claude Agent SDK and OpenAI's Codex SDK both exist to embed coding agents into other tools. What stands out is the pace: less than a month after the CLI entered beta, Meta already has an embedding path in developers' hands.
Summary
The Muse Code general release centers on three plans from 5 to 50 USD a month and an SDK in developer preview. Pay-as-you-go pricing stands at 1.25 USD per million input tokens and 4.25 USD per million output tokens, dropping to 0.10 USD and 0.20 USD for users who agree to share their data. On the feature side, inter-session messaging, Workflows and Rewind have been added. With Claude Code and Codex holding a head start, Meta is clearly betting on price and on the speed of its feature releases. Anyone choosing between them will be weighing not just the rate card but where their code and prompts end up.
