Build Independently confirmed 2 publishers 3 min read Published Sep 1, 2026, 8:20 AM PDT

Muse Code's $5 tier buys the same 10 requests per five hours that Codex Plus starts at

Meta's terminal coding agent left beta on August 31 with plans from $5 to $50. That looks like a four-to-one discount on Codex and Claude Code seats, but line up the request bands and the gap narrows to two.

The Engineer · Build desk

Send a correction

Illustration accompanying Muse Code's $5 tier buys the same 10 requests per five hours that Codex Plus starts at — Generated illustration

What happened

Meta took Muse Code out of beta on August 31st, adding cross-session messaging, workflows for coordinating teams of subagents, and monthly subscriptions that start at $5.
The three plans are Everyday Usage at $5 a month, High Usage at $15 for three times the usage, and Power Usage at $50 for ten times, with Meta estimating 10 to 50 requests per five hours on the entry tier.
The full launch also brought a rewind feature that rolls conversation and code changes back to an earlier point, plus a control room for monitoring and steering agents inside Workflows.

Why it matters

Who has to decide decision Twenty developers on a $5 seat rather than a $20 one is $3,600 a year, which is enough to justify a procurement conversation and small enough that the request band, not the sticker, decides the answer.
What it limits constraint With no published cap on how many agents a workflow runs or what governs branching, retries and spending, there is no way to bound the blast radius of a runaway workflow before signing up.
Who's exposed exposure The cheapest route through Muse Code is priced in code disclosure, which puts it out of reach for the proprietary repositories that most need the savings, according to engineering leaders who spoke to The New Stack.
What doesn't add up contradiction Meta's evidence for long-horizon durability is its own 24-hour, 1,000-tool-call kernel runs, while the one independent head-to-head found more tokens burned and weaker refactoring output, so the two point different ways.

Next story · Invest · 4d ago

Meta's Muse Wonder, at $20 to $200 a month, is framed as a test of its superintelligence team

Confirmed25 publishers

The two published request bands overlap at the bottom, and that is where the budget lever gets complicated. Meta estimates roughly 10 to 50 requests every five hours on the $5 Everyday Usage plan ^[5]. OpenAI's pricing page estimates 10 to 100 local Codex messages every five hours on the $20 Plus plan running GPT-5.6 Sol ^[10]. If your work lands at the floor of both bands, you pay a quarter as much for the same ten units of work ^[22]. If it lands at the ceiling, Plus buys twice the messages for four times the money, so the per-message discount is 2x ^[21]. Both vendors say the count moves with model, task complexity, context, reasoning and tools ^[5]^[10]. The saving you actually get depends on your repository, not on the price page.

Extrapolating the ladder gets you 30 to 150 requests per five hours at $15 ^[25] and 100 to 500 at $50 ^[23], assuming Meta's stated 3x and 10x multiples apply to the request estimate rather than to some internal token budget ^[7]. Against OpenAI's $100 Pro 5x estimate of 50 to 500 ^[10], that is the same ceiling for half the price and twice the floor ^[23]. The Claude Code comparison cannot be run at all. Anthropic publishes no absolute prompt count for Pro, Max 5x or Max 20x, states the Max allowances as multiples of an unpublished base, and applies five-hour and weekly limits on top ^[11].

The pay-as-you-go rates make the trade explicit. Standard is $1.25 per million input tokens and $4.25 per million output; Contributor is $0.10 and $0.20 ^[12]. That is 12.5x cheaper on input and 21.25x cheaper on output ^[24], and the price of admission is opting in to Meta using your prompts and completions ^[13]. For the subscription saving to survive contact with real work, Muse Code's token consumption on your tasks has to stay inside whatever headroom the rate discount gives you, and that is measurable in an afternoon on one repository.

On mechanism, the coordination features are narrower than the word "workflow" suggests. Cross-session messaging lets separate jobs exchange messages while they run, and Meta has not claimed sessions share complete transcripts or working directories, so treat it as messaging rather than synchronization ^[15]. The durable part is older: a local append-only event log of model calls, tool runs, approvals and edits, which Meta says lets an interrupted session resume where it stopped ^[17]. The cookbook documents replaying sessions in CI and isolating subagents in separate Git worktrees ^[18], which is the kind of plumbing that decides whether parallel agents corrupt each other's checkouts. Workflows ship with a control room, but Meta has not published a cap on agent count, branching, retries or spending ^[16]^[6], so there is no way to know the size of a workflow before you run it.

Adoption cost is mostly patience. The plans arrive as a staged rollout ^[19] and the SDK is a developer preview ^[3], so a migration date is not yours to set yet.

What to watch

Whether Meta publishes workflow limits on agent count, branching, retries and spending when the SDK leaves developer preview.
Whether Anthropic answers with absolute Claude Code prompt counts for Pro and Max, which would turn the seat comparison into arithmetic.
Whether the staged rollout reaches organisation accounts with training-on-prompts off by default at the subscription tiers.

Clarity's read

What the record supports and how the coverage leans. The claims behind it follow.

Reality

Evidence57
Adoption21
Hype gap+27
Incentives71
Confidence63

Why these scores

Claim ledger

Ranked by verification strength, evidence, and original report placement.

[1]
Meta took Muse Code out of beta on August 31st, adding cross-session messaging, workflows for coordinating teams of subagents, and monthly subscriptions starting at $5.
Reported SupportedView cited source
[2]
The release came 26 days after Meta introduced Muse Code as a beta product, and in a post on X Meta said the monthly plans were beginning to roll out on August 31st.
Reported SupportedView cited source
[3]
Meta put a Muse Code SDK into developer preview, giving developers an early route to build around the coding agent's runtime.
Reported SupportedView cited source
[4]
Muse Code runs in the terminal and uses Meta's Muse Spark 1.2 model to plan changes, edit code and validate its work across large repositories; Meta launched the model and agent together on August 5th.
Reported SupportedView cited source
[5]
Meta says the $5 plan typically provides around 10 to 50 requests every five hours, although the amount varies depending on the complexity of the work.
Reported SupportedView cited source
[6]
For the full launch Meta added inter-session messaging, Workflows that orchestrate teams of specialized agents with a dedicated Workflows control room, and a rewind feature that rolls conversation and code changes back to an earlier point in a session.
Reported SupportedView cited source
[7]
The Everyday Usage plan costs $5 per month, High Usage costs $15 and promises three times as much usage, and the $50 Power Usage plan offers 10 times as much.
Reported SupportedView cited source
[8]
Anthropic includes Claude Code with its $20-a-month Pro plan, while its $100 Max 5x and $200 Max 20x plans provide five and 20 times Pro's per-session usage respectively.
Reported SupportedView cited source
[9]
OpenAI includes Codex with the $20 Plus plan, while its $100 and $200 Pro tiers provide five and 20 times the usage of Plus respectively.
Reported SupportedView cited source
[10]
OpenAI's pricing page estimates, for GPT-5.6 Sol, 10 to 100 local Codex messages every five hours on Plus, 50 to 500 on Pro 5x and 200 to 2,000 on Pro 20x, and says actual usage varies with factors including the model, task complexity, context, reasoning and tools.
Reported SupportedView cited source
[11]
Anthropic does not currently publish an absolute number of Claude Code prompts for Pro, Max 5x or Max 20x, instead expressing the two Max allowances as multiples of Pro, and it also applies five-hour session limits and weekly limits.
Reported SupportedView cited source
[12]
At launch Muse Code offered two pay-as-you-go tiers: a Standard tier at $1.25 per million input tokens and $4.25 per million output tokens, and a Contributor tier at $0.10 and $0.20 respectively.
Reported SupportedView cited source
[13]
Contributor users had to opt in to having their prompts and completions used to help improve Meta's products, a trade-off several engineering leaders told The New Stack would rule it out for their proprietary code.
Reported SupportedView cited source
[14]
The New Stack's Jessica Wachtel ran Muse Code and Claude Code through the same three coding tasks; Muse Code cost much less but consumed substantially more tokens, produced a weaker result on a refactoring task and left dead code behind on another.
Reported SupportedView cited source
[15]
Meta has not claimed that Muse Code sessions share complete transcripts or working directories, so cross-session communication should be understood as messaging rather than automatic synchronization.
Reported SupportedView cited source
[16]
Meta's announcement does not specify how many agents a workflow can operate or what controls developers have over branching, retries and spending.
Reported SupportedView cited source
[17]
Muse Code is built around a local append-only event log that records model calls, tool runs, approvals and edits, and according to Meta that design lets an interrupted session resume from the point where it stopped.
Reported SupportedView cited source
[18]
The public Muse Code cookbook documents recipes for replaying sessions in continuous integration, isolating subagents in separate Git worktrees, scheduling recurring work and running side conversations outside the main transcript.
Reported SupportedView cited source
[19]
Meta described the new plans as a rollout, meaning availability may reach accounts in stages.
Reported SupportedView cited source
[20]
Meta reported testing the Muse Spark 1.2 and Muse Code pairing on GPU-kernel optimization runs lasting as long as 24 hours and spanning more than 1,000 tool calls, which remain Meta's internal evaluations rather than independent evidence.
Reported SupportedView cited source
[21]
At the ceiling of each estimate, Codex Plus quotes 100 messages against Muse Code's 50 for four times the monthly price, which works out to twice the cost per message rather than four times.
Derived Supported
[22]
At the floor of each published estimate, the $5 Muse Code plan and the $20 Codex Plus plan both quote 10 requests or messages per five hours, so Muse Code is one quarter the price for the same floor count.
Derived Supported
[23]
Applying the $50 Power Usage plan's stated 10x multiple to Meta's request estimate implies 100 to 500 requests every five hours, the same ceiling as OpenAI's $100 Pro 5x estimate of 50 to 500 and twice its floor.
Derived Supported
[24]
Muse Code's Contributor pay-as-you-go rate is 12.5 times cheaper than Standard on input tokens and 21.25 times cheaper on output tokens.
Derived Supported
[25]
Applying the $15 High Usage plan's stated 3x multiple to Meta's request estimate implies 30 to 150 requests every five hours.
Derived Supported
[26]
Twenty developers on a $5 entry plan instead of a $20 entry plan is a $300 monthly difference, or $3,600 a year.
Derived Supported

Sources

2 independent publishers whose own reporting we read for this story.

runtimewire.com
1 article · August 31, 2026
Meta takes Muse Code out of beta and adds $5 monthly plans
thenewstack.io
1 article · September 1, 2026
Meta’s Claude Code rival exits beta with three new subscription tiers — and it’s pushing hard on price

Clarity analysis

How the coverage leans and what the record actually supports. Read the methodology

Perspective Coverage

How much of the coverage speaks to builders, operators and investors. It measures emphasis, not whether the story is right.

Builder

A builder gets two halves of one product from two places. Runtime Wire has the mechanics — an append-only log of model calls, tool runs, approvals and edits that lets a killed session pick up where it stopped, cookbook recipes for replaying sessions in CI and isolating subagents in Git worktrees, and an SDK preview that turns the agent into something your own code can start and steer. The New Stack has the part that decides whether any of it is worth adopting: on the same three tasks, cheaper per token did not mean cheaper per finished job, and one refactor came back weaker with dead code left behind. Neither outlet can tell you how wide a Workflow can fan out, or what happens when a run goes sideways at $5 a month.

Operator

The allowance is quoted in requests per five hours, not seats per month, and that is the whole operator problem. Meta's 10-to-50 band carries a complexity caveat, the upper tiers are published as multiples rather than counts, and Anthropic stacks a weekly limit on top of its own five-hour window — so no one running a team can predict when work stops. The cheapest token rates still require opting prompts and completions into improving Meta's products, which engineering leaders told The New Stack rules it out for proprietary code. Add staged availability and an unspecified set of controls over branching, retries and spend, and the pilot you can plan today is small by necessity.

Investor

Meta priced entry at a quarter of what Anthropic and OpenAI charge, and the interesting number is not the discount but where it stops. Line up the published bands and $5 buys the same floor as $20 Codex Plus but half the ceiling, which is two times cheaper per message, not four. Twenty developers moved off a $20 entry plan saves $3,600 a year — real, and small enough that it is a wedge, not a business. The revenue question sits underneath: Contributor rates are 12.5 to 21 times below Standard in exchange for training data, and there are no usage, seat or conversion figures anywhere in this reporting to show whether cheap seats are filling.

Reality

Evidence, adoption, hype gap, incentives and confidence, each read independently from the current record on the scale shown beneath it. How these are measured.

Evidence57: 0100
One announcement, one hands-on run
Strip out Meta's developer blog, its X posts and the rivals' pricing pages and almost nothing is left standing. The prices, the 10-to-50 request band, the resume-from-interruption behaviour and the 24-hour kernel runs are all the vendor describing itself, relayed accurately by two outlets. The one measurement anybody took is The New Stack's three-task run, which is first-hand but unreplicated and cuts against the sales pitch.
Adoption21: 0100
Day-one, no users counted
Everything observable here happened on the vendor's side of the transaction: a version out of beta, a price sheet, an SDK preview. No deployment, no seat count, no named team running Workflows in anger, and the plans themselves are still reaching accounts in stages. The closest thing to usage evidence is one reporter's three tasks.
Hype gap+27: −100 understated0 aligned+100 overstated
The discount halves on inspection
A four-to-one price gap is the number everyone repeats, including a machine learning engineer calling the tiers refreshingly simple. Run the published bands and it is four-to-one only at the floor: at the ceiling, $5 buys 50 where $20 buys 100, which is two times cheaper per message. Then the one independent test finds the cheap agent burning substantially more tokens and shipping weaker work, so even the two-times figure flatters it on cost per finished job.
Incentives71: 0100
Cheap seats with data attached
Meta's cheapest token rates are 12.5 to 21 times below its own Standard prices in exchange for permission to train on prompts and completions — the discount is the product, and Msty AI's co-founder says as much when he calls the pricing unbelievably low and names Meta itself as the concern. Note also who is speaking: the outlet doing the price comparison is citing its own reporter's benchmark, and a rival tooling founder is quoted assessing a competitor's launch. Runtime Wire's account runs on a vendor blog post and vendor X posts.
Confidence63: 0100
Facts firm, meaning unsettled
On the checkable particulars the two accounts agree and neither contradicts the other, so dates, prices and feature names can be taken as reported. What is not settled is what the prices buy: request bands with complexity caveats on one side, undisclosed prompt counts on the other, and admitted gaps in what Meta has said about workflow scale and spending controls. Confident about the rate card, unconfident about the comparison it invites.

Topics and entities

Topics

AI Subscription Pricing
AI Coding Agents
Developer tooling economics
Training-data consent as discount
Agent Orchestration and Subagents